1 /*
   2  * Copyright (c) 2007, 2015, 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/castnode.hpp"
  31 #include "opto/convertnode.hpp"
  32 #include "opto/divnode.hpp"
  33 #include "opto/matcher.hpp"
  34 #include "opto/memnode.hpp"
  35 #include "opto/mulnode.hpp"
  36 #include "opto/opcodes.hpp"
  37 #include "opto/opaquenode.hpp"
  38 #include "opto/superword.hpp"
  39 #include "opto/vectornode.hpp"
  40 
  41 //
  42 //                  S U P E R W O R D   T R A N S F O R M
  43 //=============================================================================
  44 
  45 //------------------------------SuperWord---------------------------
  46 SuperWord::SuperWord(PhaseIdealLoop* phase) :
  47   _phase(phase),
  48   _igvn(phase->_igvn),
  49   _arena(phase->C->comp_arena()),
  50   _packset(arena(), 8,  0, NULL),         // packs for the current block
  51   _bb_idx(arena(), (int)(1.10 * phase->C->unique()), 0, 0), // node idx to index in bb
  52   _block(arena(), 8,  0, NULL),           // nodes in current block
  53   _data_entry(arena(), 8,  0, NULL),      // nodes with all inputs from outside
  54   _mem_slice_head(arena(), 8,  0, NULL),  // memory slice heads
  55   _mem_slice_tail(arena(), 8,  0, NULL),  // memory slice tails
  56   _node_info(arena(), 8,  0, SWNodeInfo::initial), // info needed per node
  57   _clone_map(phase->C->clone_map()),      // map of nodes created in cloning
  58   _align_to_ref(NULL),                    // memory reference to align vectors to
  59   _disjoint_ptrs(arena(), 8,  0, OrderedPair::initial), // runtime disambiguated pointer pairs
  60   _dg(_arena),                            // dependence graph
  61   _visited(arena()),                      // visited node set
  62   _post_visited(arena()),                 // post visited node set
  63   _n_idx_list(arena(), 8),                // scratch list of (node,index) pairs
  64   _stk(arena(), 8, 0, NULL),              // scratch stack of nodes
  65   _nlist(arena(), 8, 0, NULL),            // scratch list of nodes
  66   _lpt(NULL),                             // loop tree node
  67   _lp(NULL),                              // LoopNode
  68   _bb(NULL),                              // basic block
  69   _iv(NULL),                              // induction var
  70   _race_possible(false),                  // cases where SDMU is true
  71   _num_work_vecs(0),                      // amount of vector work we have
  72   _num_reductions(0),                     // amount of reduction work we have
  73   _do_vector_loop(phase->C->do_vector_loop()),  // whether to do vectorization/simd style
  74   _ii_first(-1),                          // first loop generation index - only if do_vector_loop()
  75   _ii_last(-1),                           // last loop generation index - only if do_vector_loop()
  76   _ii_order(arena(), 8, 0, 0),
  77   _vector_loop_debug(phase->C->has_method() && phase->C->method_has_option("VectorizeDebug"))
  78 {}
  79 
  80 //------------------------------transform_loop---------------------------
  81 void SuperWord::transform_loop(IdealLoopTree* lpt) {
  82   assert(UseSuperWord, "should be");
  83   // Do vectors exist on this architecture?
  84   if (Matcher::vector_width_in_bytes(T_BYTE) < 2) return;
  85 
  86   assert(lpt->_head->is_CountedLoop(), "must be");
  87   CountedLoopNode *cl = lpt->_head->as_CountedLoop();
  88 
  89   if (!cl->is_valid_counted_loop()) return; // skip malformed counted loop
  90 
  91   if (!cl->is_main_loop() ) return; // skip normal, pre, and post loops
  92 
  93   // Check for no control flow in body (other than exit)
  94   Node *cl_exit = cl->loopexit();
  95   if (cl_exit->in(0) != lpt->_head) return;
  96 
  97   // Make sure the are no extra control users of the loop backedge
  98   if (cl->back_control()->outcnt() != 1) {
  99     return;
 100   }
 101 
 102   // Check for pre-loop ending with CountedLoopEnd(Bool(Cmp(x,Opaque1(limit))))
 103   CountedLoopEndNode* pre_end = get_pre_loop_end(cl);
 104   if (pre_end == NULL) return;
 105   Node *pre_opaq1 = pre_end->limit();
 106   if (pre_opaq1->Opcode() != Op_Opaque1) return;
 107 
 108   init(); // initialize data structures
 109 
 110   set_lpt(lpt);
 111   set_lp(cl);
 112 
 113   // For now, define one block which is the entire loop body
 114   set_bb(cl);
 115 
 116   assert(_packset.length() == 0, "packset must be empty");
 117   SLP_extract();
 118 }
 119 
 120 //------------------------------SLP_extract---------------------------
 121 // Extract the superword level parallelism
 122 //
 123 // 1) A reverse post-order of nodes in the block is constructed.  By scanning
 124 //    this list from first to last, all definitions are visited before their uses.
 125 //
 126 // 2) A point-to-point dependence graph is constructed between memory references.
 127 //    This simplies the upcoming "independence" checker.
 128 //
 129 // 3) The maximum depth in the node graph from the beginning of the block
 130 //    to each node is computed.  This is used to prune the graph search
 131 //    in the independence checker.
 132 //
 133 // 4) For integer types, the necessary bit width is propagated backwards
 134 //    from stores to allow packed operations on byte, char, and short
 135 //    integers.  This reverses the promotion to type "int" that javac
 136 //    did for operations like: char c1,c2,c3;  c1 = c2 + c3.
 137 //
 138 // 5) One of the memory references is picked to be an aligned vector reference.
 139 //    The pre-loop trip count is adjusted to align this reference in the
 140 //    unrolled body.
 141 //
 142 // 6) The initial set of pack pairs is seeded with memory references.
 143 //
 144 // 7) The set of pack pairs is extended by following use->def and def->use links.
 145 //
 146 // 8) The pairs are combined into vector sized packs.
 147 //
 148 // 9) Reorder the memory slices to co-locate members of the memory packs.
 149 //
 150 // 10) Generate ideal vector nodes for the final set of packs and where necessary,
 151 //    inserting scalar promotion, vector creation from multiple scalars, and
 152 //    extraction of scalar values from vectors.
 153 //
 154 void SuperWord::SLP_extract() {
 155 
 156 #ifndef PRODUCT
 157   if (_do_vector_loop && TraceSuperWord) {
 158     tty->print("SuperWord::SLP_extract\n");
 159     tty->print("input loop\n");
 160     _lpt->dump_head();
 161     _lpt->dump();
 162     for (uint i = 0; i < _lpt->_body.size(); i++) {
 163       _lpt->_body.at(i)->dump();
 164     }
 165   }
 166 #endif
 167   // Ready the block
 168   if (!construct_bb()) {
 169     return; // Exit if no interesting nodes or complex graph.
 170   }
 171   // build    _dg, _disjoint_ptrs
 172   dependence_graph();
 173 
 174   // compute function depth(Node*)
 175   compute_max_depth();
 176 
 177   if (_do_vector_loop) {
 178     if (mark_generations() != -1) {
 179       hoist_loads_in_graph(); // this only rebuild the graph; all basic structs need rebuild explicitly
 180 
 181       if (!construct_bb()) {
 182         return; // Exit if no interesting nodes or complex graph.
 183       }
 184       dependence_graph();
 185       compute_max_depth();
 186     }
 187 
 188 #ifndef PRODUCT
 189     if (TraceSuperWord) {
 190       tty->print_cr("\nSuperWord::_do_vector_loop: graph after hoist_loads_in_graph");
 191       _lpt->dump_head();
 192       for (int j = 0; j < _block.length(); j++) {
 193         Node* n = _block.at(j);
 194         int d = depth(n);
 195         for (int i = 0;  i < d; i++) tty->print("%s", "  ");
 196         tty->print("%d :", d);
 197         n->dump();
 198       }
 199     }
 200 #endif
 201   }
 202 
 203   compute_vector_element_type();
 204 
 205   // Attempt vectorization
 206 
 207   find_adjacent_refs();
 208 
 209   extend_packlist();
 210 
 211   if (_do_vector_loop) {
 212     if (_packset.length() == 0) {
 213 #ifndef PRODUCT
 214       if (TraceSuperWord) {
 215         tty->print_cr("\nSuperWord::_do_vector_loop DFA could not build packset, now trying to build anyway");
 216       }
 217 #endif
 218       pack_parallel();
 219     }
 220   }
 221 
 222   combine_packs();
 223 
 224   construct_my_pack_map();
 225 
 226   filter_packs();
 227 
 228   schedule();
 229 
 230   output();
 231 }
 232 
 233 //------------------------------find_adjacent_refs---------------------------
 234 // Find the adjacent memory references and create pack pairs for them.
 235 // This is the initial set of packs that will then be extended by
 236 // following use->def and def->use links.  The align positions are
 237 // assigned relative to the reference "align_to_ref"
 238 void SuperWord::find_adjacent_refs() {
 239   // Get list of memory operations
 240   Node_List memops;
 241   for (int i = 0; i < _block.length(); i++) {
 242     Node* n = _block.at(i);
 243     if (n->is_Mem() && !n->is_LoadStore() && in_bb(n) &&
 244         is_java_primitive(n->as_Mem()->memory_type())) {
 245       int align = memory_alignment(n->as_Mem(), 0);
 246       if (align != bottom_align) {
 247         memops.push(n);
 248       }
 249     }
 250   }
 251 
 252   Node_List align_to_refs;
 253   int best_iv_adjustment = 0;
 254   MemNode* best_align_to_mem_ref = NULL;
 255 
 256   while (memops.size() != 0) {
 257     // Find a memory reference to align to.
 258     MemNode* mem_ref = find_align_to_ref(memops);
 259     if (mem_ref == NULL) break;
 260     align_to_refs.push(mem_ref);
 261     int iv_adjustment = get_iv_adjustment(mem_ref);
 262 
 263     if (best_align_to_mem_ref == NULL) {
 264       // Set memory reference which is the best from all memory operations
 265       // to be used for alignment. The pre-loop trip count is modified to align
 266       // this reference to a vector-aligned address.
 267       best_align_to_mem_ref = mem_ref;
 268       best_iv_adjustment = iv_adjustment;
 269     }
 270 
 271     SWPointer align_to_ref_p(mem_ref, this);
 272     // Set alignment relative to "align_to_ref" for all related memory operations.
 273     for (int i = memops.size() - 1; i >= 0; i--) {
 274       MemNode* s = memops.at(i)->as_Mem();
 275       if (isomorphic(s, mem_ref)) {
 276         SWPointer p2(s, this);
 277         if (p2.comparable(align_to_ref_p)) {
 278           int align = memory_alignment(s, iv_adjustment);
 279           set_alignment(s, align);
 280         }
 281       }
 282     }
 283 
 284     // Create initial pack pairs of memory operations for which
 285     // alignment is set and vectors will be aligned.
 286     bool create_pack = true;
 287     if (memory_alignment(mem_ref, best_iv_adjustment) == 0 || _do_vector_loop) {
 288       if (!Matcher::misaligned_vectors_ok()) {
 289         int vw = vector_width(mem_ref);
 290         int vw_best = vector_width(best_align_to_mem_ref);
 291         if (vw > vw_best) {
 292           // Do not vectorize a memory access with more elements per vector
 293           // if unaligned memory access is not allowed because number of
 294           // iterations in pre-loop will be not enough to align it.
 295           create_pack = false;
 296         } else {
 297           SWPointer p2(best_align_to_mem_ref, this);
 298           if (align_to_ref_p.invar() != p2.invar()) {
 299             // Do not vectorize memory accesses with different invariants
 300             // if unaligned memory accesses are not allowed.
 301             create_pack = false;
 302           }
 303         }
 304       }
 305     } else {
 306       if (same_velt_type(mem_ref, best_align_to_mem_ref)) {
 307         // Can't allow vectorization of unaligned memory accesses with the
 308         // same type since it could be overlapped accesses to the same array.
 309         create_pack = false;
 310       } else {
 311         // Allow independent (different type) unaligned memory operations
 312         // if HW supports them.
 313         if (!Matcher::misaligned_vectors_ok()) {
 314           create_pack = false;
 315         } else {
 316           // Check if packs of the same memory type but
 317           // with a different alignment were created before.
 318           for (uint i = 0; i < align_to_refs.size(); i++) {
 319             MemNode* mr = align_to_refs.at(i)->as_Mem();
 320             if (same_velt_type(mr, mem_ref) &&
 321                 memory_alignment(mr, iv_adjustment) != 0)
 322               create_pack = false;
 323           }
 324         }
 325       }
 326     }
 327     if (create_pack) {
 328       for (uint i = 0; i < memops.size(); i++) {
 329         Node* s1 = memops.at(i);
 330         int align = alignment(s1);
 331         if (align == top_align) continue;
 332         for (uint j = 0; j < memops.size(); j++) {
 333           Node* s2 = memops.at(j);
 334           if (alignment(s2) == top_align) continue;
 335           if (s1 != s2 && are_adjacent_refs(s1, s2)) {
 336             if (stmts_can_pack(s1, s2, align)) {
 337               Node_List* pair = new Node_List();
 338               pair->push(s1);
 339               pair->push(s2);
 340               if (!_do_vector_loop || _clone_map.idx(s1->_idx) == _clone_map.idx(s2->_idx)) {
 341                 _packset.append(pair);
 342               }
 343             }
 344           }
 345         }
 346       }
 347     } else { // Don't create unaligned pack
 348       // First, remove remaining memory ops of the same type from the list.
 349       for (int i = memops.size() - 1; i >= 0; i--) {
 350         MemNode* s = memops.at(i)->as_Mem();
 351         if (same_velt_type(s, mem_ref)) {
 352           memops.remove(i);
 353         }
 354       }
 355 
 356       // Second, remove already constructed packs of the same type.
 357       for (int i = _packset.length() - 1; i >= 0; i--) {
 358         Node_List* p = _packset.at(i);
 359         MemNode* s = p->at(0)->as_Mem();
 360         if (same_velt_type(s, mem_ref)) {
 361           remove_pack_at(i);
 362         }
 363       }
 364 
 365       // If needed find the best memory reference for loop alignment again.
 366       if (same_velt_type(mem_ref, best_align_to_mem_ref)) {
 367         // Put memory ops from remaining packs back on memops list for
 368         // the best alignment search.
 369         uint orig_msize = memops.size();
 370         for (int i = 0; i < _packset.length(); i++) {
 371           Node_List* p = _packset.at(i);
 372           MemNode* s = p->at(0)->as_Mem();
 373           assert(!same_velt_type(s, mem_ref), "sanity");
 374           memops.push(s);
 375         }
 376         MemNode* best_align_to_mem_ref = find_align_to_ref(memops);
 377         if (best_align_to_mem_ref == NULL) break;
 378         best_iv_adjustment = get_iv_adjustment(best_align_to_mem_ref);
 379         // Restore list.
 380         while (memops.size() > orig_msize)
 381           (void)memops.pop();
 382       }
 383     } // unaligned memory accesses
 384 
 385     // Remove used mem nodes.
 386     for (int i = memops.size() - 1; i >= 0; i--) {
 387       MemNode* m = memops.at(i)->as_Mem();
 388       if (alignment(m) != top_align) {
 389         memops.remove(i);
 390       }
 391     }
 392 
 393   } // while (memops.size() != 0
 394   set_align_to_ref(best_align_to_mem_ref);
 395 
 396 #ifndef PRODUCT
 397   if (TraceSuperWord) {
 398     tty->print_cr("\nAfter find_adjacent_refs");
 399     print_packset();
 400   }
 401 #endif
 402 }
 403 
 404 //------------------------------find_align_to_ref---------------------------
 405 // Find a memory reference to align the loop induction variable to.
 406 // Looks first at stores then at loads, looking for a memory reference
 407 // with the largest number of references similar to it.
 408 MemNode* SuperWord::find_align_to_ref(Node_List &memops) {
 409   GrowableArray<int> cmp_ct(arena(), memops.size(), memops.size(), 0);
 410 
 411   // Count number of comparable memory ops
 412   for (uint i = 0; i < memops.size(); i++) {
 413     MemNode* s1 = memops.at(i)->as_Mem();
 414     SWPointer p1(s1, this);
 415     // Discard if pre loop can't align this reference
 416     if (!ref_is_alignable(p1)) {
 417       *cmp_ct.adr_at(i) = 0;
 418       continue;
 419     }
 420     for (uint j = i+1; j < memops.size(); j++) {
 421       MemNode* s2 = memops.at(j)->as_Mem();
 422       if (isomorphic(s1, s2)) {
 423         SWPointer p2(s2, this);
 424         if (p1.comparable(p2)) {
 425           (*cmp_ct.adr_at(i))++;
 426           (*cmp_ct.adr_at(j))++;
 427         }
 428       }
 429     }
 430   }
 431 
 432   // Find Store (or Load) with the greatest number of "comparable" references,
 433   // biggest vector size, smallest data size and smallest iv offset.
 434   int max_ct        = 0;
 435   int max_vw        = 0;
 436   int max_idx       = -1;
 437   int min_size      = max_jint;
 438   int min_iv_offset = max_jint;
 439   for (uint j = 0; j < memops.size(); j++) {
 440     MemNode* s = memops.at(j)->as_Mem();
 441     if (s->is_Store()) {
 442       int vw = vector_width_in_bytes(s);
 443       assert(vw > 1, "sanity");
 444       SWPointer p(s, this);
 445       if (cmp_ct.at(j) >  max_ct ||
 446           cmp_ct.at(j) == max_ct &&
 447             (vw >  max_vw ||
 448              vw == max_vw &&
 449               (data_size(s) <  min_size ||
 450                data_size(s) == min_size &&
 451                  (p.offset_in_bytes() < min_iv_offset)))) {
 452         max_ct = cmp_ct.at(j);
 453         max_vw = vw;
 454         max_idx = j;
 455         min_size = data_size(s);
 456         min_iv_offset = p.offset_in_bytes();
 457       }
 458     }
 459   }
 460   // If no stores, look at loads
 461   if (max_ct == 0) {
 462     for (uint j = 0; j < memops.size(); j++) {
 463       MemNode* s = memops.at(j)->as_Mem();
 464       if (s->is_Load()) {
 465         int vw = vector_width_in_bytes(s);
 466         assert(vw > 1, "sanity");
 467         SWPointer p(s, this);
 468         if (cmp_ct.at(j) >  max_ct ||
 469             cmp_ct.at(j) == max_ct &&
 470               (vw >  max_vw ||
 471                vw == max_vw &&
 472                 (data_size(s) <  min_size ||
 473                  data_size(s) == min_size &&
 474                    (p.offset_in_bytes() < min_iv_offset)))) {
 475           max_ct = cmp_ct.at(j);
 476           max_vw = vw;
 477           max_idx = j;
 478           min_size = data_size(s);
 479           min_iv_offset = p.offset_in_bytes();
 480         }
 481       }
 482     }
 483   }
 484 
 485 #ifdef ASSERT
 486   if (TraceSuperWord && Verbose) {
 487     tty->print_cr("\nVector memops after find_align_to_ref");
 488     for (uint i = 0; i < memops.size(); i++) {
 489       MemNode* s = memops.at(i)->as_Mem();
 490       s->dump();
 491     }
 492   }
 493 #endif
 494 
 495   if (max_ct > 0) {
 496 #ifdef ASSERT
 497     if (TraceSuperWord) {
 498       tty->print("\nVector align to node: ");
 499       memops.at(max_idx)->as_Mem()->dump();
 500     }
 501 #endif
 502     return memops.at(max_idx)->as_Mem();
 503   }
 504   return NULL;
 505 }
 506 
 507 //------------------------------ref_is_alignable---------------------------
 508 // Can the preloop align the reference to position zero in the vector?
 509 bool SuperWord::ref_is_alignable(SWPointer& p) {
 510   if (!p.has_iv()) {
 511     return true;   // no induction variable
 512   }
 513   CountedLoopEndNode* pre_end = get_pre_loop_end(lp()->as_CountedLoop());
 514   assert(pre_end != NULL, "we must have a correct pre-loop");
 515   assert(pre_end->stride_is_con(), "pre loop stride is constant");
 516   int preloop_stride = pre_end->stride_con();
 517 
 518   int span = preloop_stride * p.scale_in_bytes();
 519   int mem_size = p.memory_size();
 520   int offset   = p.offset_in_bytes();
 521   // Stride one accesses are alignable if offset is aligned to memory operation size.
 522   // Offset can be unaligned when UseUnalignedAccesses is used.
 523   if (ABS(span) == mem_size && (ABS(offset) % mem_size) == 0) {
 524     return true;
 525   }
 526   // If the initial offset from start of the object is computable,
 527   // check if the pre-loop can align the final offset accordingly.
 528   //
 529   // In other words: Can we find an i such that the offset
 530   // after i pre-loop iterations is aligned to vw?
 531   //   (init_offset + pre_loop) % vw == 0              (1)
 532   // where
 533   //   pre_loop = i * span
 534   // is the number of bytes added to the offset by i pre-loop iterations.
 535   //
 536   // For this to hold we need pre_loop to increase init_offset by
 537   //   pre_loop = vw - (init_offset % vw)
 538   //
 539   // This is only possible if pre_loop is divisible by span because each
 540   // pre-loop iteration increases the initial offset by 'span' bytes:
 541   //   (vw - (init_offset % vw)) % span == 0
 542   //
 543   int vw = vector_width_in_bytes(p.mem());
 544   assert(vw > 1, "sanity");
 545   Node* init_nd = pre_end->init_trip();
 546   if (init_nd->is_Con() && p.invar() == NULL) {
 547     int init = init_nd->bottom_type()->is_int()->get_con();
 548     int init_offset = init * p.scale_in_bytes() + offset;
 549     assert(init_offset >= 0, "positive offset from object start");
 550     if (vw % span == 0) {
 551       // If vm is a multiple of span, we use formula (1).
 552       if (span > 0) {
 553         return (vw - (init_offset % vw)) % span == 0;
 554       } else {
 555         assert(span < 0, "nonzero stride * scale");
 556         return (init_offset % vw) % -span == 0;
 557       }
 558     } else if (span % vw == 0) {
 559       // If span is a multiple of vw, we can simplify formula (1) to:
 560       //   (init_offset + i * span) % vw == 0
 561       //     =>
 562       //   (init_offset % vw) + ((i * span) % vw) == 0
 563       //     =>
 564       //   init_offset % vw == 0
 565       //
 566       // Because we add a multiple of vw to the initial offset, the final
 567       // offset is a multiple of vw if and only if init_offset is a multiple.
 568       //
 569       return (init_offset % vw) == 0;
 570     }
 571   }
 572   return false;
 573 }
 574 
 575 //---------------------------get_iv_adjustment---------------------------
 576 // Calculate loop's iv adjustment for this memory ops.
 577 int SuperWord::get_iv_adjustment(MemNode* mem_ref) {
 578   SWPointer align_to_ref_p(mem_ref, this);
 579   int offset = align_to_ref_p.offset_in_bytes();
 580   int scale  = align_to_ref_p.scale_in_bytes();
 581   int elt_size = align_to_ref_p.memory_size();
 582   int vw       = vector_width_in_bytes(mem_ref);
 583   assert(vw > 1, "sanity");
 584   int iv_adjustment;
 585   if (scale != 0) {
 586     int stride_sign = (scale * iv_stride()) > 0 ? 1 : -1;
 587     // At least one iteration is executed in pre-loop by default. As result
 588     // several iterations are needed to align memory operations in main-loop even
 589     // if offset is 0.
 590     int iv_adjustment_in_bytes = (stride_sign * vw - (offset % vw));
 591     assert(((ABS(iv_adjustment_in_bytes) % elt_size) == 0),
 592            err_msg_res("(%d) should be divisible by (%d)", iv_adjustment_in_bytes, elt_size));
 593     iv_adjustment = iv_adjustment_in_bytes/elt_size;
 594   } else {
 595     // This memory op is not dependent on iv (scale == 0)
 596     iv_adjustment = 0;
 597   }
 598 
 599 #ifndef PRODUCT
 600   if (TraceSuperWord)
 601     tty->print_cr("\noffset = %d iv_adjust = %d elt_size = %d scale = %d iv_stride = %d vect_size %d",
 602                   offset, iv_adjustment, elt_size, scale, iv_stride(), vw);
 603 #endif
 604   return iv_adjustment;
 605 }
 606 
 607 //---------------------------dependence_graph---------------------------
 608 // Construct dependency graph.
 609 // Add dependence edges to load/store nodes for memory dependence
 610 //    A.out()->DependNode.in(1) and DependNode.out()->B.prec(x)
 611 void SuperWord::dependence_graph() {
 612   // First, assign a dependence node to each memory node
 613   for (int i = 0; i < _block.length(); i++ ) {
 614     Node *n = _block.at(i);
 615     if (n->is_Mem() || n->is_Phi() && n->bottom_type() == Type::MEMORY) {
 616       _dg.make_node(n);
 617     }
 618   }
 619 
 620   // For each memory slice, create the dependences
 621   for (int i = 0; i < _mem_slice_head.length(); i++) {
 622     Node* n      = _mem_slice_head.at(i);
 623     Node* n_tail = _mem_slice_tail.at(i);
 624 
 625     // Get slice in predecessor order (last is first)
 626     mem_slice_preds(n_tail, n, _nlist);
 627 
 628 #ifndef PRODUCT
 629     if(TraceSuperWord && Verbose) {
 630       tty->print_cr("SuperWord::dependence_graph: built a new mem slice");
 631       for (int j = _nlist.length() - 1; j >= 0 ; j--) {
 632         _nlist.at(j)->dump();
 633       }
 634     }
 635 #endif
 636     // Make the slice dependent on the root
 637     DepMem* slice = _dg.dep(n);
 638     _dg.make_edge(_dg.root(), slice);
 639 
 640     // Create a sink for the slice
 641     DepMem* slice_sink = _dg.make_node(NULL);
 642     _dg.make_edge(slice_sink, _dg.tail());
 643 
 644     // Now visit each pair of memory ops, creating the edges
 645     for (int j = _nlist.length() - 1; j >= 0 ; j--) {
 646       Node* s1 = _nlist.at(j);
 647 
 648       // If no dependency yet, use slice
 649       if (_dg.dep(s1)->in_cnt() == 0) {
 650         _dg.make_edge(slice, s1);
 651       }
 652       SWPointer p1(s1->as_Mem(), this);
 653       bool sink_dependent = true;
 654       for (int k = j - 1; k >= 0; k--) {
 655         Node* s2 = _nlist.at(k);
 656         if (s1->is_Load() && s2->is_Load())
 657           continue;
 658         SWPointer p2(s2->as_Mem(), this);
 659 
 660         int cmp = p1.cmp(p2);
 661         if (SuperWordRTDepCheck &&
 662             p1.base() != p2.base() && p1.valid() && p2.valid()) {
 663           // Create a runtime check to disambiguate
 664           OrderedPair pp(p1.base(), p2.base());
 665           _disjoint_ptrs.append_if_missing(pp);
 666         } else if (!SWPointer::not_equal(cmp)) {
 667           // Possibly same address
 668           _dg.make_edge(s1, s2);
 669           sink_dependent = false;
 670         }
 671       }
 672       if (sink_dependent) {
 673         _dg.make_edge(s1, slice_sink);
 674       }
 675     }
 676 #ifndef PRODUCT
 677     if (TraceSuperWord) {
 678       tty->print_cr("\nDependence graph for slice: %d", n->_idx);
 679       for (int q = 0; q < _nlist.length(); q++) {
 680         _dg.print(_nlist.at(q));
 681       }
 682       tty->cr();
 683     }
 684 #endif
 685     _nlist.clear();
 686   }
 687 
 688 #ifndef PRODUCT
 689   if (TraceSuperWord) {
 690     tty->print_cr("\ndisjoint_ptrs: %s", _disjoint_ptrs.length() > 0 ? "" : "NONE");
 691     for (int r = 0; r < _disjoint_ptrs.length(); r++) {
 692       _disjoint_ptrs.at(r).print();
 693       tty->cr();
 694     }
 695     tty->cr();
 696   }
 697 #endif
 698 }
 699 
 700 //---------------------------mem_slice_preds---------------------------
 701 // Return a memory slice (node list) in predecessor order starting at "start"
 702 void SuperWord::mem_slice_preds(Node* start, Node* stop, GrowableArray<Node*> &preds) {
 703   assert(preds.length() == 0, "start empty");
 704   Node* n = start;
 705   Node* prev = NULL;
 706   while (true) {
 707     assert(in_bb(n), "must be in block");
 708     for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
 709       Node* out = n->fast_out(i);
 710       if (out->is_Load()) {
 711         if (in_bb(out)) {
 712           preds.push(out);
 713         }
 714       } else {
 715         // FIXME
 716         if (out->is_MergeMem() && !in_bb(out)) {
 717           // Either unrolling is causing a memory edge not to disappear,
 718           // or need to run igvn.optimize() again before SLP
 719         } else if (out->is_Phi() && out->bottom_type() == Type::MEMORY && !in_bb(out)) {
 720           // Ditto.  Not sure what else to check further.
 721         } else if (out->Opcode() == Op_StoreCM && out->in(MemNode::OopStore) == n) {
 722           // StoreCM has an input edge used as a precedence edge.
 723           // Maybe an issue when oop stores are vectorized.
 724         } else {
 725           assert(out == prev || prev == NULL, "no branches off of store slice");
 726         }
 727       }
 728     }
 729     if (n == stop) break;
 730     preds.push(n);
 731     prev = n;
 732     assert(n->is_Mem(), err_msg_res("unexpected node %s", n->Name()));
 733     n = n->in(MemNode::Memory);
 734   }
 735 }
 736 
 737 //------------------------------stmts_can_pack---------------------------
 738 // Can s1 and s2 be in a pack with s1 immediately preceding s2 and
 739 // s1 aligned at "align"
 740 bool SuperWord::stmts_can_pack(Node* s1, Node* s2, int align) {
 741 
 742   // Do not use superword for non-primitives
 743   BasicType bt1 = velt_basic_type(s1);
 744   BasicType bt2 = velt_basic_type(s2);
 745   if(!is_java_primitive(bt1) || !is_java_primitive(bt2))
 746     return false;
 747   if (Matcher::max_vector_size(bt1) < 2) {
 748     return false; // No vectors for this type
 749   }
 750 
 751   if (isomorphic(s1, s2)) {
 752     if (independent(s1, s2) || reduction(s1, s2)) {
 753       if (!exists_at(s1, 0) && !exists_at(s2, 1)) {
 754         if (!s1->is_Mem() || are_adjacent_refs(s1, s2)) {
 755           int s1_align = alignment(s1);
 756           int s2_align = alignment(s2);
 757           if (s1_align == top_align || s1_align == align) {
 758             if (s2_align == top_align || s2_align == align + data_size(s1)) {
 759               return true;
 760             }
 761           }
 762         }
 763       }
 764     }
 765   }
 766   return false;
 767 }
 768 
 769 //------------------------------exists_at---------------------------
 770 // Does s exist in a pack at position pos?
 771 bool SuperWord::exists_at(Node* s, uint pos) {
 772   for (int i = 0; i < _packset.length(); i++) {
 773     Node_List* p = _packset.at(i);
 774     if (p->at(pos) == s) {
 775       return true;
 776     }
 777   }
 778   return false;
 779 }
 780 
 781 //------------------------------are_adjacent_refs---------------------------
 782 // Is s1 immediately before s2 in memory?
 783 bool SuperWord::are_adjacent_refs(Node* s1, Node* s2) {
 784   if (!s1->is_Mem() || !s2->is_Mem()) return false;
 785   if (!in_bb(s1)    || !in_bb(s2))    return false;
 786 
 787   // Do not use superword for non-primitives
 788   if (!is_java_primitive(s1->as_Mem()->memory_type()) ||
 789       !is_java_primitive(s2->as_Mem()->memory_type())) {
 790     return false;
 791   }
 792 
 793   // FIXME - co_locate_pack fails on Stores in different mem-slices, so
 794   // only pack memops that are in the same alias set until that's fixed.
 795   if (_phase->C->get_alias_index(s1->as_Mem()->adr_type()) !=
 796       _phase->C->get_alias_index(s2->as_Mem()->adr_type()))
 797     return false;
 798   SWPointer p1(s1->as_Mem(), this);
 799   SWPointer p2(s2->as_Mem(), this);
 800   if (p1.base() != p2.base() || !p1.comparable(p2)) return false;
 801   int diff = p2.offset_in_bytes() - p1.offset_in_bytes();
 802   return diff == data_size(s1);
 803 }
 804 
 805 //------------------------------isomorphic---------------------------
 806 // Are s1 and s2 similar?
 807 bool SuperWord::isomorphic(Node* s1, Node* s2) {
 808   if (s1->Opcode() != s2->Opcode()) return false;
 809   if (s1->req() != s2->req()) return false;
 810   if (s1->in(0) != s2->in(0)) return false;
 811   if (!same_velt_type(s1, s2)) return false;
 812   return true;
 813 }
 814 
 815 //------------------------------independent---------------------------
 816 // Is there no data path from s1 to s2 or s2 to s1?
 817 bool SuperWord::independent(Node* s1, Node* s2) {
 818   //  assert(s1->Opcode() == s2->Opcode(), "check isomorphic first");
 819   int d1 = depth(s1);
 820   int d2 = depth(s2);
 821   if (d1 == d2) return s1 != s2;
 822   Node* deep    = d1 > d2 ? s1 : s2;
 823   Node* shallow = d1 > d2 ? s2 : s1;
 824 
 825   visited_clear();
 826 
 827   return independent_path(shallow, deep);
 828 }
 829 
 830 //------------------------------reduction---------------------------
 831 // Is there a data path between s1 and s2 and the nodes reductions?
 832 bool SuperWord::reduction(Node* s1, Node* s2) {
 833   bool retValue = false;
 834   int d1 = depth(s1);
 835   int d2 = depth(s2);
 836   if (d1 + 1 == d2) {
 837     if (s1->is_reduction() && s2->is_reduction()) {
 838       // This is an ordered set, so s1 should define s2
 839       for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
 840         Node* t1 = s1->fast_out(i);
 841         if (t1 == s2) {
 842           // both nodes are reductions and connected
 843           retValue = true;
 844         }
 845       }
 846     }
 847   }
 848 
 849   return retValue;
 850 }
 851 
 852 //------------------------------independent_path------------------------------
 853 // Helper for independent
 854 bool SuperWord::independent_path(Node* shallow, Node* deep, uint dp) {
 855   if (dp >= 1000) return false; // stop deep recursion
 856   visited_set(deep);
 857   int shal_depth = depth(shallow);
 858   assert(shal_depth <= depth(deep), "must be");
 859   for (DepPreds preds(deep, _dg); !preds.done(); preds.next()) {
 860     Node* pred = preds.current();
 861     if (in_bb(pred) && !visited_test(pred)) {
 862       if (shallow == pred) {
 863         return false;
 864       }
 865       if (shal_depth < depth(pred) && !independent_path(shallow, pred, dp+1)) {
 866         return false;
 867       }
 868     }
 869   }
 870   return true;
 871 }
 872 
 873 //------------------------------set_alignment---------------------------
 874 void SuperWord::set_alignment(Node* s1, Node* s2, int align) {
 875   set_alignment(s1, align);
 876   if (align == top_align || align == bottom_align) {
 877     set_alignment(s2, align);
 878   } else {
 879     set_alignment(s2, align + data_size(s1));
 880   }
 881 }
 882 
 883 //------------------------------data_size---------------------------
 884 int SuperWord::data_size(Node* s) {
 885   int bsize = type2aelembytes(velt_basic_type(s));
 886   assert(bsize != 0, "valid size");
 887   return bsize;
 888 }
 889 
 890 //------------------------------extend_packlist---------------------------
 891 // Extend packset by following use->def and def->use links from pack members.
 892 void SuperWord::extend_packlist() {
 893   bool changed;
 894   do {
 895     packset_sort(_packset.length());
 896     changed = false;
 897     for (int i = 0; i < _packset.length(); i++) {
 898       Node_List* p = _packset.at(i);
 899       changed |= follow_use_defs(p);
 900       changed |= follow_def_uses(p);
 901     }
 902   } while (changed);
 903 
 904   if (_race_possible) {
 905     for (int i = 0; i < _packset.length(); i++) {
 906       Node_List* p = _packset.at(i);
 907       order_def_uses(p);
 908     }
 909   }
 910 
 911 #ifndef PRODUCT
 912   if (TraceSuperWord) {
 913     tty->print_cr("\nAfter extend_packlist");
 914     print_packset();
 915   }
 916 #endif
 917 }
 918 
 919 //------------------------------follow_use_defs---------------------------
 920 // Extend the packset by visiting operand definitions of nodes in pack p
 921 bool SuperWord::follow_use_defs(Node_List* p) {
 922   assert(p->size() == 2, "just checking");
 923   Node* s1 = p->at(0);
 924   Node* s2 = p->at(1);
 925   assert(s1->req() == s2->req(), "just checking");
 926   assert(alignment(s1) + data_size(s1) == alignment(s2), "just checking");
 927 
 928   if (s1->is_Load()) return false;
 929 
 930   int align = alignment(s1);
 931   bool changed = false;
 932   int start = s1->is_Store() ? MemNode::ValueIn   : 1;
 933   int end   = s1->is_Store() ? MemNode::ValueIn+1 : s1->req();
 934   for (int j = start; j < end; j++) {
 935     Node* t1 = s1->in(j);
 936     Node* t2 = s2->in(j);
 937     if (!in_bb(t1) || !in_bb(t2))
 938       continue;
 939     if (stmts_can_pack(t1, t2, align)) {
 940       if (est_savings(t1, t2) >= 0) {
 941         Node_List* pair = new Node_List();
 942         pair->push(t1);
 943         pair->push(t2);
 944         _packset.append(pair);
 945         set_alignment(t1, t2, align);
 946         changed = true;
 947       }
 948     }
 949   }
 950   return changed;
 951 }
 952 
 953 //------------------------------follow_def_uses---------------------------
 954 // Extend the packset by visiting uses of nodes in pack p
 955 bool SuperWord::follow_def_uses(Node_List* p) {
 956   bool changed = false;
 957   Node* s1 = p->at(0);
 958   Node* s2 = p->at(1);
 959   assert(p->size() == 2, "just checking");
 960   assert(s1->req() == s2->req(), "just checking");
 961   assert(alignment(s1) + data_size(s1) == alignment(s2), "just checking");
 962 
 963   if (s1->is_Store()) return false;
 964 
 965   int align = alignment(s1);
 966   int savings = -1;
 967   int num_s1_uses = 0;
 968   Node* u1 = NULL;
 969   Node* u2 = NULL;
 970   for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
 971     Node* t1 = s1->fast_out(i);
 972     num_s1_uses++;
 973     if (!in_bb(t1)) continue;
 974     for (DUIterator_Fast jmax, j = s2->fast_outs(jmax); j < jmax; j++) {
 975       Node* t2 = s2->fast_out(j);
 976       if (!in_bb(t2)) continue;
 977       if (!opnd_positions_match(s1, t1, s2, t2))
 978         continue;
 979       if (stmts_can_pack(t1, t2, align)) {
 980         int my_savings = est_savings(t1, t2);
 981         if (my_savings > savings) {
 982           savings = my_savings;
 983           u1 = t1;
 984           u2 = t2;
 985         }
 986       }
 987     }
 988   }
 989   if (num_s1_uses > 1) {
 990     _race_possible = true;
 991   }
 992   if (savings >= 0) {
 993     Node_List* pair = new Node_List();
 994     pair->push(u1);
 995     pair->push(u2);
 996     _packset.append(pair);
 997     set_alignment(u1, u2, align);
 998     changed = true;
 999   }
1000   return changed;
1001 }
1002 
1003 //------------------------------order_def_uses---------------------------
1004 // For extended packsets, ordinally arrange uses packset by major component
1005 void SuperWord::order_def_uses(Node_List* p) {
1006   Node* s1 = p->at(0);
1007 
1008   if (s1->is_Store()) return;
1009 
1010   // reductions are always managed beforehand
1011   if (s1->is_reduction()) return;
1012 
1013   for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
1014     Node* t1 = s1->fast_out(i);
1015 
1016     // Only allow operand swap on commuting operations
1017     if (!t1->is_Add() && !t1->is_Mul()) {
1018       break;
1019     }
1020 
1021     // Now find t1's packset
1022     Node_List* p2 = NULL;
1023     for (int j = 0; j < _packset.length(); j++) {
1024       p2 = _packset.at(j);
1025       Node* first = p2->at(0);
1026       if (t1 == first) {
1027         break;
1028       }
1029       p2 = NULL;
1030     }
1031     // Arrange all sub components by the major component
1032     if (p2 != NULL) {
1033       for (uint j = 1; j < p->size(); j++) {
1034         Node* d1 = p->at(j);
1035         Node* u1 = p2->at(j);
1036         opnd_positions_match(s1, t1, d1, u1);
1037       }
1038     }
1039   }
1040 }
1041 
1042 //---------------------------opnd_positions_match-------------------------
1043 // Is the use of d1 in u1 at the same operand position as d2 in u2?
1044 bool SuperWord::opnd_positions_match(Node* d1, Node* u1, Node* d2, Node* u2) {
1045   // check reductions to see if they are marshalled to represent the reduction
1046   // operator in a specified opnd
1047   if (u1->is_reduction() && u2->is_reduction()) {
1048     // ensure reductions have phis and reduction definitions feeding the 1st operand
1049     Node* first = u1->in(2);
1050     if (first->is_Phi() || first->is_reduction()) {
1051       u1->swap_edges(1, 2);
1052     }
1053     // ensure reductions have phis and reduction definitions feeding the 1st operand
1054     first = u2->in(2);
1055     if (first->is_Phi() || first->is_reduction()) {
1056       u2->swap_edges(1, 2);
1057     }
1058     return true;
1059   }
1060 
1061   uint ct = u1->req();
1062   if (ct != u2->req()) return false;
1063   uint i1 = 0;
1064   uint i2 = 0;
1065   do {
1066     for (i1++; i1 < ct; i1++) if (u1->in(i1) == d1) break;
1067     for (i2++; i2 < ct; i2++) if (u2->in(i2) == d2) break;
1068     if (i1 != i2) {
1069       if ((i1 == (3-i2)) && (u2->is_Add() || u2->is_Mul())) {
1070         // Further analysis relies on operands position matching.
1071         u2->swap_edges(i1, i2);
1072       } else {
1073         return false;
1074       }
1075     }
1076   } while (i1 < ct);
1077   return true;
1078 }
1079 
1080 //------------------------------est_savings---------------------------
1081 // Estimate the savings from executing s1 and s2 as a pack
1082 int SuperWord::est_savings(Node* s1, Node* s2) {
1083   int save_in = 2 - 1; // 2 operations per instruction in packed form
1084 
1085   // inputs
1086   for (uint i = 1; i < s1->req(); i++) {
1087     Node* x1 = s1->in(i);
1088     Node* x2 = s2->in(i);
1089     if (x1 != x2) {
1090       if (are_adjacent_refs(x1, x2)) {
1091         save_in += adjacent_profit(x1, x2);
1092       } else if (!in_packset(x1, x2)) {
1093         save_in -= pack_cost(2);
1094       } else {
1095         save_in += unpack_cost(2);
1096       }
1097     }
1098   }
1099 
1100   // uses of result
1101   uint ct = 0;
1102   int save_use = 0;
1103   for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
1104     Node* s1_use = s1->fast_out(i);
1105     for (int j = 0; j < _packset.length(); j++) {
1106       Node_List* p = _packset.at(j);
1107       if (p->at(0) == s1_use) {
1108         for (DUIterator_Fast kmax, k = s2->fast_outs(kmax); k < kmax; k++) {
1109           Node* s2_use = s2->fast_out(k);
1110           if (p->at(p->size()-1) == s2_use) {
1111             ct++;
1112             if (are_adjacent_refs(s1_use, s2_use)) {
1113               save_use += adjacent_profit(s1_use, s2_use);
1114             }
1115           }
1116         }
1117       }
1118     }
1119   }
1120 
1121   if (ct < s1->outcnt()) save_use += unpack_cost(1);
1122   if (ct < s2->outcnt()) save_use += unpack_cost(1);
1123 
1124   return MAX2(save_in, save_use);
1125 }
1126 
1127 //------------------------------costs---------------------------
1128 int SuperWord::adjacent_profit(Node* s1, Node* s2) { return 2; }
1129 int SuperWord::pack_cost(int ct)   { return ct; }
1130 int SuperWord::unpack_cost(int ct) { return ct; }
1131 
1132 //------------------------------combine_packs---------------------------
1133 // Combine packs A and B with A.last == B.first into A.first..,A.last,B.second,..B.last
1134 void SuperWord::combine_packs() {
1135   bool changed = true;
1136   // Combine packs regardless max vector size.
1137   while (changed) {
1138     changed = false;
1139     for (int i = 0; i < _packset.length(); i++) {
1140       Node_List* p1 = _packset.at(i);
1141       if (p1 == NULL) continue;
1142       // Because of sorting we can start at i + 1
1143       for (int j = i + 1; j < _packset.length(); j++) {
1144         Node_List* p2 = _packset.at(j);
1145         if (p2 == NULL) continue;
1146         if (i == j) continue;
1147         if (p1->at(p1->size()-1) == p2->at(0)) {
1148           for (uint k = 1; k < p2->size(); k++) {
1149             p1->push(p2->at(k));
1150           }
1151           _packset.at_put(j, NULL);
1152           changed = true;
1153         }
1154       }
1155     }
1156   }
1157 
1158   // Split packs which have size greater then max vector size.
1159   for (int i = 0; i < _packset.length(); i++) {
1160     Node_List* p1 = _packset.at(i);
1161     if (p1 != NULL) {
1162       BasicType bt = velt_basic_type(p1->at(0));
1163       uint max_vlen = Matcher::max_vector_size(bt); // Max elements in vector
1164       assert(is_power_of_2(max_vlen), "sanity");
1165       uint psize = p1->size();
1166       if (!is_power_of_2(psize)) {
1167         // Skip pack which can't be vector.
1168         // case1: for(...) { a[i] = i; }    elements values are different (i+x)
1169         // case2: for(...) { a[i] = b[i+1]; }  can't align both, load and store
1170         _packset.at_put(i, NULL);
1171         continue;
1172       }
1173       if (psize > max_vlen) {
1174         Node_List* pack = new Node_List();
1175         for (uint j = 0; j < psize; j++) {
1176           pack->push(p1->at(j));
1177           if (pack->size() >= max_vlen) {
1178             assert(is_power_of_2(pack->size()), "sanity");
1179             _packset.append(pack);
1180             pack = new Node_List();
1181           }
1182         }
1183         _packset.at_put(i, NULL);
1184       }
1185     }
1186   }
1187 
1188   // Compress list.
1189   for (int i = _packset.length() - 1; i >= 0; i--) {
1190     Node_List* p1 = _packset.at(i);
1191     if (p1 == NULL) {
1192       _packset.remove_at(i);
1193     }
1194   }
1195 
1196 #ifndef PRODUCT
1197   if (TraceSuperWord) {
1198     tty->print_cr("\nAfter combine_packs");
1199     print_packset();
1200   }
1201 #endif
1202 }
1203 
1204 //-----------------------------construct_my_pack_map--------------------------
1205 // Construct the map from nodes to packs.  Only valid after the
1206 // point where a node is only in one pack (after combine_packs).
1207 void SuperWord::construct_my_pack_map() {
1208   Node_List* rslt = NULL;
1209   for (int i = 0; i < _packset.length(); i++) {
1210     Node_List* p = _packset.at(i);
1211     for (uint j = 0; j < p->size(); j++) {
1212       Node* s = p->at(j);
1213       assert(my_pack(s) == NULL, "only in one pack");
1214       set_my_pack(s, p);
1215     }
1216   }
1217 }
1218 
1219 //------------------------------filter_packs---------------------------
1220 // Remove packs that are not implemented or not profitable.
1221 void SuperWord::filter_packs() {
1222   // Remove packs that are not implemented
1223   for (int i = _packset.length() - 1; i >= 0; i--) {
1224     Node_List* pk = _packset.at(i);
1225     bool impl = implemented(pk);
1226     if (!impl) {
1227 #ifndef PRODUCT
1228       if (TraceSuperWord && Verbose) {
1229         tty->print_cr("Unimplemented");
1230         pk->at(0)->dump();
1231       }
1232 #endif
1233       remove_pack_at(i);
1234     }
1235     Node *n = pk->at(0);
1236     if (n->is_reduction()) {
1237       _num_reductions++;
1238     } else {
1239       _num_work_vecs++;
1240     }
1241   }
1242 
1243   // Remove packs that are not profitable
1244   bool changed;
1245   do {
1246     changed = false;
1247     for (int i = _packset.length() - 1; i >= 0; i--) {
1248       Node_List* pk = _packset.at(i);
1249       bool prof = profitable(pk);
1250       if (!prof) {
1251 #ifndef PRODUCT
1252         if (TraceSuperWord && Verbose) {
1253           tty->print_cr("Unprofitable");
1254           pk->at(0)->dump();
1255         }
1256 #endif
1257         remove_pack_at(i);
1258         changed = true;
1259       }
1260     }
1261   } while (changed);
1262 
1263 #ifndef PRODUCT
1264   if (TraceSuperWord) {
1265     tty->print_cr("\nAfter filter_packs");
1266     print_packset();
1267     tty->cr();
1268   }
1269 #endif
1270 }
1271 
1272 //------------------------------implemented---------------------------
1273 // Can code be generated for pack p?
1274 bool SuperWord::implemented(Node_List* p) {
1275   bool retValue = false;
1276   Node* p0 = p->at(0);
1277   if (p0 != NULL) {
1278     int opc = p0->Opcode();
1279     uint size = p->size();
1280     if (p0->is_reduction()) {
1281       const Type *arith_type = p0->bottom_type();
1282       // Length 2 reductions of INT/LONG do not offer performance benefits
1283       if (((arith_type->basic_type() == T_INT) || (arith_type->basic_type() == T_LONG)) && (size == 2)) {
1284         retValue = false;
1285       } else {
1286         retValue = ReductionNode::implemented(opc, size, arith_type->basic_type());
1287       }
1288     } else {
1289       retValue = VectorNode::implemented(opc, size, velt_basic_type(p0));
1290     }
1291   }
1292   return retValue;
1293 }
1294 
1295 //------------------------------same_inputs--------------------------
1296 // For pack p, are all idx operands the same?
1297 static bool same_inputs(Node_List* p, int idx) {
1298   Node* p0 = p->at(0);
1299   uint vlen = p->size();
1300   Node* p0_def = p0->in(idx);
1301   for (uint i = 1; i < vlen; i++) {
1302     Node* pi = p->at(i);
1303     Node* pi_def = pi->in(idx);
1304     if (p0_def != pi_def)
1305       return false;
1306   }
1307   return true;
1308 }
1309 
1310 //------------------------------profitable---------------------------
1311 // For pack p, are all operands and all uses (with in the block) vector?
1312 bool SuperWord::profitable(Node_List* p) {
1313   Node* p0 = p->at(0);
1314   uint start, end;
1315   VectorNode::vector_operands(p0, &start, &end);
1316 
1317   // Return false if some inputs are not vectors or vectors with different
1318   // size or alignment.
1319   // Also, for now, return false if not scalar promotion case when inputs are
1320   // the same. Later, implement PackNode and allow differing, non-vector inputs
1321   // (maybe just the ones from outside the block.)
1322   for (uint i = start; i < end; i++) {
1323     if (!is_vector_use(p0, i))
1324       return false;
1325   }
1326   // Check if reductions are connected
1327   if (p0->is_reduction()) {
1328     Node* second_in = p0->in(2);
1329     Node_List* second_pk = my_pack(second_in);
1330     if ((second_pk == NULL) || (_num_work_vecs == _num_reductions)) {
1331       // Remove reduction flag if no parent pack or if not enough work
1332       // to cover reduction expansion overhead
1333       p0->remove_flag(Node::Flag_is_reduction);
1334       return false;
1335     } else if (second_pk->size() != p->size()) {
1336       return false;
1337     }
1338   }
1339   if (VectorNode::is_shift(p0)) {
1340     // For now, return false if shift count is vector or not scalar promotion
1341     // case (different shift counts) because it is not supported yet.
1342     Node* cnt = p0->in(2);
1343     Node_List* cnt_pk = my_pack(cnt);
1344     if (cnt_pk != NULL)
1345       return false;
1346     if (!same_inputs(p, 2))
1347       return false;
1348   }
1349   if (!p0->is_Store()) {
1350     // For now, return false if not all uses are vector.
1351     // Later, implement ExtractNode and allow non-vector uses (maybe
1352     // just the ones outside the block.)
1353     for (uint i = 0; i < p->size(); i++) {
1354       Node* def = p->at(i);
1355       for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
1356         Node* use = def->fast_out(j);
1357         for (uint k = 0; k < use->req(); k++) {
1358           Node* n = use->in(k);
1359           if (def == n) {
1360             // reductions can be loop carried dependences
1361             if (def->is_reduction() && use->is_Phi())
1362               continue;
1363             if (!is_vector_use(use, k)) {
1364               return false;
1365             }
1366           }
1367         }
1368       }
1369     }
1370   }
1371   return true;
1372 }
1373 
1374 //------------------------------schedule---------------------------
1375 // Adjust the memory graph for the packed operations
1376 void SuperWord::schedule() {
1377 
1378   // Co-locate in the memory graph the members of each memory pack
1379   for (int i = 0; i < _packset.length(); i++) {
1380     co_locate_pack(_packset.at(i));
1381   }
1382 }
1383 
1384 //-------------------------------remove_and_insert-------------------
1385 // Remove "current" from its current position in the memory graph and insert
1386 // it after the appropriate insertion point (lip or uip).
1387 void SuperWord::remove_and_insert(MemNode *current, MemNode *prev, MemNode *lip,
1388                                   Node *uip, Unique_Node_List &sched_before) {
1389   Node* my_mem = current->in(MemNode::Memory);
1390   bool sched_up = sched_before.member(current);
1391 
1392   // remove current_store from its current position in the memmory graph
1393   for (DUIterator i = current->outs(); current->has_out(i); i++) {
1394     Node* use = current->out(i);
1395     if (use->is_Mem()) {
1396       assert(use->in(MemNode::Memory) == current, "must be");
1397       if (use == prev) { // connect prev to my_mem
1398           _igvn.replace_input_of(use, MemNode::Memory, my_mem);
1399           --i; //deleted this edge; rescan position
1400       } else if (sched_before.member(use)) {
1401         if (!sched_up) { // Will be moved together with current
1402           _igvn.replace_input_of(use, MemNode::Memory, uip);
1403           --i; //deleted this edge; rescan position
1404         }
1405       } else {
1406         if (sched_up) { // Will be moved together with current
1407           _igvn.replace_input_of(use, MemNode::Memory, lip);
1408           --i; //deleted this edge; rescan position
1409         }
1410       }
1411     }
1412   }
1413 
1414   Node *insert_pt =  sched_up ?  uip : lip;
1415 
1416   // all uses of insert_pt's memory state should use current's instead
1417   for (DUIterator i = insert_pt->outs(); insert_pt->has_out(i); i++) {
1418     Node* use = insert_pt->out(i);
1419     if (use->is_Mem()) {
1420       assert(use->in(MemNode::Memory) == insert_pt, "must be");
1421       _igvn.replace_input_of(use, MemNode::Memory, current);
1422       --i; //deleted this edge; rescan position
1423     } else if (!sched_up && use->is_Phi() && use->bottom_type() == Type::MEMORY) {
1424       uint pos; //lip (lower insert point) must be the last one in the memory slice
1425       for (pos=1; pos < use->req(); pos++) {
1426         if (use->in(pos) == insert_pt) break;
1427       }
1428       _igvn.replace_input_of(use, pos, current);
1429       --i;
1430     }
1431   }
1432 
1433   //connect current to insert_pt
1434   _igvn.replace_input_of(current, MemNode::Memory, insert_pt);
1435 }
1436 
1437 //------------------------------co_locate_pack----------------------------------
1438 // To schedule a store pack, we need to move any sandwiched memory ops either before
1439 // or after the pack, based upon dependence information:
1440 // (1) If any store in the pack depends on the sandwiched memory op, the
1441 //     sandwiched memory op must be scheduled BEFORE the pack;
1442 // (2) If a sandwiched memory op depends on any store in the pack, the
1443 //     sandwiched memory op must be scheduled AFTER the pack;
1444 // (3) If a sandwiched memory op (say, memA) depends on another sandwiched
1445 //     memory op (say memB), memB must be scheduled before memA. So, if memA is
1446 //     scheduled before the pack, memB must also be scheduled before the pack;
1447 // (4) If there is no dependence restriction for a sandwiched memory op, we simply
1448 //     schedule this store AFTER the pack
1449 // (5) We know there is no dependence cycle, so there in no other case;
1450 // (6) Finally, all memory ops in another single pack should be moved in the same direction.
1451 //
1452 // To schedule a load pack, we use the memory state of either the first or the last load in
1453 // the pack, based on the dependence constraint.
1454 void SuperWord::co_locate_pack(Node_List* pk) {
1455   if (pk->at(0)->is_Store()) {
1456     MemNode* first     = executed_first(pk)->as_Mem();
1457     MemNode* last      = executed_last(pk)->as_Mem();
1458     Unique_Node_List schedule_before_pack;
1459     Unique_Node_List memops;
1460 
1461     MemNode* current   = last->in(MemNode::Memory)->as_Mem();
1462     MemNode* previous  = last;
1463     while (true) {
1464       assert(in_bb(current), "stay in block");
1465       memops.push(previous);
1466       for (DUIterator i = current->outs(); current->has_out(i); i++) {
1467         Node* use = current->out(i);
1468         if (use->is_Mem() && use != previous)
1469           memops.push(use);
1470       }
1471       if (current == first) break;
1472       previous = current;
1473       current  = current->in(MemNode::Memory)->as_Mem();
1474     }
1475 
1476     // determine which memory operations should be scheduled before the pack
1477     for (uint i = 1; i < memops.size(); i++) {
1478       Node *s1 = memops.at(i);
1479       if (!in_pack(s1, pk) && !schedule_before_pack.member(s1)) {
1480         for (uint j = 0; j< i; j++) {
1481           Node *s2 = memops.at(j);
1482           if (!independent(s1, s2)) {
1483             if (in_pack(s2, pk) || schedule_before_pack.member(s2)) {
1484               schedule_before_pack.push(s1); // s1 must be scheduled before
1485               Node_List* mem_pk = my_pack(s1);
1486               if (mem_pk != NULL) {
1487                 for (uint ii = 0; ii < mem_pk->size(); ii++) {
1488                   Node* s = mem_pk->at(ii);  // follow partner
1489                   if (memops.member(s) && !schedule_before_pack.member(s))
1490                     schedule_before_pack.push(s);
1491                 }
1492               }
1493               break;
1494             }
1495           }
1496         }
1497       }
1498     }
1499 
1500     Node*    upper_insert_pt = first->in(MemNode::Memory);
1501     // Following code moves loads connected to upper_insert_pt below aliased stores.
1502     // Collect such loads here and reconnect them back to upper_insert_pt later.
1503     memops.clear();
1504     for (DUIterator i = upper_insert_pt->outs(); upper_insert_pt->has_out(i); i++) {
1505       Node* use = upper_insert_pt->out(i);
1506       if (use->is_Mem() && !use->is_Store()) {
1507         memops.push(use);
1508       }
1509     }
1510 
1511     MemNode* lower_insert_pt = last;
1512     previous                 = last; //previous store in pk
1513     current                  = last->in(MemNode::Memory)->as_Mem();
1514 
1515     // start scheduling from "last" to "first"
1516     while (true) {
1517       assert(in_bb(current), "stay in block");
1518       assert(in_pack(previous, pk), "previous stays in pack");
1519       Node* my_mem = current->in(MemNode::Memory);
1520 
1521       if (in_pack(current, pk)) {
1522         // Forward users of my memory state (except "previous) to my input memory state
1523         for (DUIterator i = current->outs(); current->has_out(i); i++) {
1524           Node* use = current->out(i);
1525           if (use->is_Mem() && use != previous) {
1526             assert(use->in(MemNode::Memory) == current, "must be");
1527             if (schedule_before_pack.member(use)) {
1528               _igvn.replace_input_of(use, MemNode::Memory, upper_insert_pt);
1529             } else {
1530               _igvn.replace_input_of(use, MemNode::Memory, lower_insert_pt);
1531             }
1532             --i; // deleted this edge; rescan position
1533           }
1534         }
1535         previous = current;
1536       } else { // !in_pack(current, pk) ==> a sandwiched store
1537         remove_and_insert(current, previous, lower_insert_pt, upper_insert_pt, schedule_before_pack);
1538       }
1539 
1540       if (current == first) break;
1541       current = my_mem->as_Mem();
1542     } // end while
1543 
1544     // Reconnect loads back to upper_insert_pt.
1545     for (uint i = 0; i < memops.size(); i++) {
1546       Node *ld = memops.at(i);
1547       if (ld->in(MemNode::Memory) != upper_insert_pt) {
1548         _igvn.replace_input_of(ld, MemNode::Memory, upper_insert_pt);
1549       }
1550     }
1551   } else if (pk->at(0)->is_Load()) { //load
1552     // all loads in the pack should have the same memory state. By default,
1553     // we use the memory state of the last load. However, if any load could
1554     // not be moved down due to the dependence constraint, we use the memory
1555     // state of the first load.
1556     Node* last_mem  = executed_last(pk)->in(MemNode::Memory);
1557     Node* first_mem = executed_first(pk)->in(MemNode::Memory);
1558     bool schedule_last = true;
1559     for (uint i = 0; i < pk->size(); i++) {
1560       Node* ld = pk->at(i);
1561       for (Node* current = last_mem; current != ld->in(MemNode::Memory);
1562            current=current->in(MemNode::Memory)) {
1563         assert(current != first_mem, "corrupted memory graph");
1564         if(current->is_Mem() && !independent(current, ld)){
1565           schedule_last = false; // a later store depends on this load
1566           break;
1567         }
1568       }
1569     }
1570 
1571     Node* mem_input = schedule_last ? last_mem : first_mem;
1572     _igvn.hash_delete(mem_input);
1573     // Give each load the same memory state
1574     for (uint i = 0; i < pk->size(); i++) {
1575       LoadNode* ld = pk->at(i)->as_Load();
1576       _igvn.replace_input_of(ld, MemNode::Memory, mem_input);
1577     }
1578   }
1579 }
1580 
1581 //------------------------------output---------------------------
1582 // Convert packs into vector node operations
1583 void SuperWord::output() {
1584   if (_packset.length() == 0) return;
1585 
1586 #ifndef PRODUCT
1587   if (TraceLoopOpts) {
1588     tty->print("SuperWord    ");
1589     lpt()->dump_head();
1590   }
1591 #endif
1592 
1593   // MUST ENSURE main loop's initial value is properly aligned:
1594   //  (iv_initial_value + min_iv_offset) % vector_width_in_bytes() == 0
1595 
1596   align_initial_loop_index(align_to_ref());
1597 
1598   // Insert extract (unpack) operations for scalar uses
1599   for (int i = 0; i < _packset.length(); i++) {
1600     insert_extracts(_packset.at(i));
1601   }
1602 
1603   Compile* C = _phase->C;
1604   uint max_vlen_in_bytes = 0;
1605   for (int i = 0; i < _block.length(); i++) {
1606     Node* n = _block.at(i);
1607     Node_List* p = my_pack(n);
1608     if (p && n == executed_last(p)) {
1609       uint vlen = p->size();
1610       uint vlen_in_bytes = 0;
1611       Node* vn = NULL;
1612       Node* low_adr = p->at(0);
1613       Node* first   = executed_first(p);
1614       int   opc = n->Opcode();
1615       if (n->is_Load()) {
1616         Node* ctl = n->in(MemNode::Control);
1617         Node* mem = first->in(MemNode::Memory);
1618         SWPointer p1(n->as_Mem(), this);
1619         // Identify the memory dependency for the new loadVector node by
1620         // walking up through memory chain.
1621         // This is done to give flexibility to the new loadVector node so that
1622         // it can move above independent storeVector nodes.
1623         while (mem->is_StoreVector()) {
1624           SWPointer p2(mem->as_Mem(), this);
1625           int cmp = p1.cmp(p2);
1626           if (SWPointer::not_equal(cmp) || !SWPointer::comparable(cmp)) {
1627             mem = mem->in(MemNode::Memory);
1628           } else {
1629             break; // dependent memory
1630           }
1631         }
1632         Node* adr = low_adr->in(MemNode::Address);
1633         const TypePtr* atyp = n->adr_type();
1634         vn = LoadVectorNode::make(opc, ctl, mem, adr, atyp, vlen, velt_basic_type(n), control_dependency(p));
1635         vlen_in_bytes = vn->as_LoadVector()->memory_size();
1636       } else if (n->is_Store()) {
1637         // Promote value to be stored to vector
1638         Node* val = vector_opd(p, MemNode::ValueIn);
1639         Node* ctl = n->in(MemNode::Control);
1640         Node* mem = first->in(MemNode::Memory);
1641         Node* adr = low_adr->in(MemNode::Address);
1642         const TypePtr* atyp = n->adr_type();
1643         vn = StoreVectorNode::make(opc, ctl, mem, adr, atyp, val, vlen);
1644         vlen_in_bytes = vn->as_StoreVector()->memory_size();
1645       } else if (n->req() == 3) {
1646         // Promote operands to vector
1647         Node* in1 = NULL;
1648         bool node_isa_reduction = n->is_reduction();
1649         if (node_isa_reduction) {
1650           // the input to the first reduction operation is retained
1651           in1 = low_adr->in(1);
1652         } else {
1653           in1 = vector_opd(p, 1);
1654         }
1655         Node* in2 = vector_opd(p, 2);
1656         if (VectorNode::is_invariant_vector(in1) && (node_isa_reduction == false) && (n->is_Add() || n->is_Mul())) {
1657           // Move invariant vector input into second position to avoid register spilling.
1658           Node* tmp = in1;
1659           in1 = in2;
1660           in2 = tmp;
1661         }
1662         if (node_isa_reduction) {
1663           const Type *arith_type = n->bottom_type();
1664           vn = ReductionNode::make(opc, NULL, in1, in2, arith_type->basic_type());
1665           if (in2->is_Load()) {
1666             vlen_in_bytes = in2->as_LoadVector()->memory_size();
1667           } else {
1668             vlen_in_bytes = in2->as_Vector()->length_in_bytes();
1669           }
1670         } else {
1671           vn = VectorNode::make(opc, in1, in2, vlen, velt_basic_type(n));
1672           vlen_in_bytes = vn->as_Vector()->length_in_bytes();
1673         }
1674       } else {
1675         ShouldNotReachHere();
1676       }
1677       assert(vn != NULL, "sanity");
1678       _igvn.register_new_node_with_optimizer(vn);
1679       _phase->set_ctrl(vn, _phase->get_ctrl(p->at(0)));
1680       for (uint j = 0; j < p->size(); j++) {
1681         Node* pm = p->at(j);
1682         _igvn.replace_node(pm, vn);
1683       }
1684       _igvn._worklist.push(vn);
1685 
1686       if (vlen_in_bytes > max_vlen_in_bytes) {
1687         max_vlen_in_bytes = vlen_in_bytes;
1688       }
1689 #ifdef ASSERT
1690       if (TraceNewVectors) {
1691         tty->print("new Vector node: ");
1692         vn->dump();
1693       }
1694 #endif
1695     }
1696   }
1697   C->set_max_vector_size(max_vlen_in_bytes);
1698 }
1699 
1700 //------------------------------vector_opd---------------------------
1701 // Create a vector operand for the nodes in pack p for operand: in(opd_idx)
1702 Node* SuperWord::vector_opd(Node_List* p, int opd_idx) {
1703   Node* p0 = p->at(0);
1704   uint vlen = p->size();
1705   Node* opd = p0->in(opd_idx);
1706 
1707   if (same_inputs(p, opd_idx)) {
1708     if (opd->is_Vector() || opd->is_LoadVector()) {
1709       assert(((opd_idx != 2) || !VectorNode::is_shift(p0)), "shift's count can't be vector");
1710       return opd; // input is matching vector
1711     }
1712     if ((opd_idx == 2) && VectorNode::is_shift(p0)) {
1713       Compile* C = _phase->C;
1714       Node* cnt = opd;
1715       // Vector instructions do not mask shift count, do it here.
1716       juint mask = (p0->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
1717       const TypeInt* t = opd->find_int_type();
1718       if (t != NULL && t->is_con()) {
1719         juint shift = t->get_con();
1720         if (shift > mask) { // Unsigned cmp
1721           cnt = ConNode::make(TypeInt::make(shift & mask));
1722         }
1723       } else {
1724         if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
1725           cnt = ConNode::make(TypeInt::make(mask));
1726           _igvn.register_new_node_with_optimizer(cnt);
1727           cnt = new AndINode(opd, cnt);
1728           _igvn.register_new_node_with_optimizer(cnt);
1729           _phase->set_ctrl(cnt, _phase->get_ctrl(opd));
1730         }
1731         assert(opd->bottom_type()->isa_int(), "int type only");
1732         // Move non constant shift count into vector register.
1733         cnt = VectorNode::shift_count(p0, cnt, vlen, velt_basic_type(p0));
1734       }
1735       if (cnt != opd) {
1736         _igvn.register_new_node_with_optimizer(cnt);
1737         _phase->set_ctrl(cnt, _phase->get_ctrl(opd));
1738       }
1739       return cnt;
1740     }
1741     assert(!opd->is_StoreVector(), "such vector is not expected here");
1742     // Convert scalar input to vector with the same number of elements as
1743     // p0's vector. Use p0's type because size of operand's container in
1744     // vector should match p0's size regardless operand's size.
1745     const Type* p0_t = velt_type(p0);
1746     VectorNode* vn = VectorNode::scalar2vector(opd, vlen, p0_t);
1747 
1748     _igvn.register_new_node_with_optimizer(vn);
1749     _phase->set_ctrl(vn, _phase->get_ctrl(opd));
1750 #ifdef ASSERT
1751     if (TraceNewVectors) {
1752       tty->print("new Vector node: ");
1753       vn->dump();
1754     }
1755 #endif
1756     return vn;
1757   }
1758 
1759   // Insert pack operation
1760   BasicType bt = velt_basic_type(p0);
1761   PackNode* pk = PackNode::make(opd, vlen, bt);
1762   DEBUG_ONLY( const BasicType opd_bt = opd->bottom_type()->basic_type(); )
1763 
1764   for (uint i = 1; i < vlen; i++) {
1765     Node* pi = p->at(i);
1766     Node* in = pi->in(opd_idx);
1767     assert(my_pack(in) == NULL, "Should already have been unpacked");
1768     assert(opd_bt == in->bottom_type()->basic_type(), "all same type");
1769     pk->add_opd(in);
1770   }
1771   _igvn.register_new_node_with_optimizer(pk);
1772   _phase->set_ctrl(pk, _phase->get_ctrl(opd));
1773 #ifdef ASSERT
1774   if (TraceNewVectors) {
1775     tty->print("new Vector node: ");
1776     pk->dump();
1777   }
1778 #endif
1779   return pk;
1780 }
1781 
1782 //------------------------------insert_extracts---------------------------
1783 // If a use of pack p is not a vector use, then replace the
1784 // use with an extract operation.
1785 void SuperWord::insert_extracts(Node_List* p) {
1786   if (p->at(0)->is_Store()) return;
1787   assert(_n_idx_list.is_empty(), "empty (node,index) list");
1788 
1789   // Inspect each use of each pack member.  For each use that is
1790   // not a vector use, replace the use with an extract operation.
1791 
1792   for (uint i = 0; i < p->size(); i++) {
1793     Node* def = p->at(i);
1794     for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
1795       Node* use = def->fast_out(j);
1796       for (uint k = 0; k < use->req(); k++) {
1797         Node* n = use->in(k);
1798         if (def == n) {
1799           if (!is_vector_use(use, k)) {
1800             _n_idx_list.push(use, k);
1801           }
1802         }
1803       }
1804     }
1805   }
1806 
1807   while (_n_idx_list.is_nonempty()) {
1808     Node* use = _n_idx_list.node();
1809     int   idx = _n_idx_list.index();
1810     _n_idx_list.pop();
1811     Node* def = use->in(idx);
1812 
1813     if (def->is_reduction()) continue;
1814 
1815     // Insert extract operation
1816     _igvn.hash_delete(def);
1817     int def_pos = alignment(def) / data_size(def);
1818 
1819     Node* ex = ExtractNode::make(def, def_pos, velt_basic_type(def));
1820     _igvn.register_new_node_with_optimizer(ex);
1821     _phase->set_ctrl(ex, _phase->get_ctrl(def));
1822     _igvn.replace_input_of(use, idx, ex);
1823     _igvn._worklist.push(def);
1824 
1825     bb_insert_after(ex, bb_idx(def));
1826     set_velt_type(ex, velt_type(def));
1827   }
1828 }
1829 
1830 //------------------------------is_vector_use---------------------------
1831 // Is use->in(u_idx) a vector use?
1832 bool SuperWord::is_vector_use(Node* use, int u_idx) {
1833   Node_List* u_pk = my_pack(use);
1834   if (u_pk == NULL) return false;
1835   if (use->is_reduction()) return true;
1836   Node* def = use->in(u_idx);
1837   Node_List* d_pk = my_pack(def);
1838   if (d_pk == NULL) {
1839     // check for scalar promotion
1840     Node* n = u_pk->at(0)->in(u_idx);
1841     for (uint i = 1; i < u_pk->size(); i++) {
1842       if (u_pk->at(i)->in(u_idx) != n) return false;
1843     }
1844     return true;
1845   }
1846   if (u_pk->size() != d_pk->size())
1847     return false;
1848   for (uint i = 0; i < u_pk->size(); i++) {
1849     Node* ui = u_pk->at(i);
1850     Node* di = d_pk->at(i);
1851     if (ui->in(u_idx) != di || alignment(ui) != alignment(di))
1852       return false;
1853   }
1854   return true;
1855 }
1856 
1857 //------------------------------construct_bb---------------------------
1858 // Construct reverse postorder list of block members
1859 bool SuperWord::construct_bb() {
1860   Node* entry = bb();
1861 
1862   assert(_stk.length() == 0,            "stk is empty");
1863   assert(_block.length() == 0,          "block is empty");
1864   assert(_data_entry.length() == 0,     "data_entry is empty");
1865   assert(_mem_slice_head.length() == 0, "mem_slice_head is empty");
1866   assert(_mem_slice_tail.length() == 0, "mem_slice_tail is empty");
1867 
1868   // Find non-control nodes with no inputs from within block,
1869   // create a temporary map from node _idx to bb_idx for use
1870   // by the visited and post_visited sets,
1871   // and count number of nodes in block.
1872   int bb_ct = 0;
1873   for (uint i = 0; i < lpt()->_body.size(); i++) {
1874     Node *n = lpt()->_body.at(i);
1875     set_bb_idx(n, i); // Create a temporary map
1876     if (in_bb(n)) {
1877       if (n->is_LoadStore() || n->is_MergeMem() ||
1878           (n->is_Proj() && !n->as_Proj()->is_CFG())) {
1879         // Bailout if the loop has LoadStore, MergeMem or data Proj
1880         // nodes. Superword optimization does not work with them.
1881         return false;
1882       }
1883       bb_ct++;
1884       if (!n->is_CFG()) {
1885         bool found = false;
1886         for (uint j = 0; j < n->req(); j++) {
1887           Node* def = n->in(j);
1888           if (def && in_bb(def)) {
1889             found = true;
1890             break;
1891           }
1892         }
1893         if (!found) {
1894           assert(n != entry, "can't be entry");
1895           _data_entry.push(n);
1896         }
1897       }
1898     }
1899   }
1900 
1901   // Find memory slices (head and tail)
1902   for (DUIterator_Fast imax, i = lp()->fast_outs(imax); i < imax; i++) {
1903     Node *n = lp()->fast_out(i);
1904     if (in_bb(n) && (n->is_Phi() && n->bottom_type() == Type::MEMORY)) {
1905       Node* n_tail  = n->in(LoopNode::LoopBackControl);
1906       if (n_tail != n->in(LoopNode::EntryControl)) {
1907         if (!n_tail->is_Mem()) {
1908           assert(n_tail->is_Mem(), err_msg_res("unexpected node for memory slice: %s", n_tail->Name()));
1909           return false; // Bailout
1910         }
1911         _mem_slice_head.push(n);
1912         _mem_slice_tail.push(n_tail);
1913       }
1914     }
1915   }
1916 
1917   // Create an RPO list of nodes in block
1918 
1919   visited_clear();
1920   post_visited_clear();
1921 
1922   // Push all non-control nodes with no inputs from within block, then control entry
1923   for (int j = 0; j < _data_entry.length(); j++) {
1924     Node* n = _data_entry.at(j);
1925     visited_set(n);
1926     _stk.push(n);
1927   }
1928   visited_set(entry);
1929   _stk.push(entry);
1930 
1931   // Do a depth first walk over out edges
1932   int rpo_idx = bb_ct - 1;
1933   int size;
1934   int reduction_uses = 0;
1935   while ((size = _stk.length()) > 0) {
1936     Node* n = _stk.top(); // Leave node on stack
1937     if (!visited_test_set(n)) {
1938       // forward arc in graph
1939     } else if (!post_visited_test(n)) {
1940       // cross or back arc
1941       for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1942         Node *use = n->fast_out(i);
1943         if (in_bb(use) && !visited_test(use) &&
1944             // Don't go around backedge
1945             (!use->is_Phi() || n == entry)) {
1946           if (use->is_reduction()) {
1947             // First see if we can map the reduction on the given system we are on, then
1948             // make a data entry operation for each reduction we see.
1949             BasicType bt = use->bottom_type()->basic_type();
1950             if (ReductionNode::implemented(use->Opcode(), Matcher::min_vector_size(bt), bt)) {
1951               reduction_uses++;
1952             }
1953           }
1954           _stk.push(use);
1955         }
1956       }
1957       if (_stk.length() == size) {
1958         // There were no additional uses, post visit node now
1959         _stk.pop(); // Remove node from stack
1960         assert(rpo_idx >= 0, "");
1961         _block.at_put_grow(rpo_idx, n);
1962         rpo_idx--;
1963         post_visited_set(n);
1964         assert(rpo_idx >= 0 || _stk.is_empty(), "");
1965       }
1966     } else {
1967       _stk.pop(); // Remove post-visited node from stack
1968     }
1969   }
1970 
1971   // Create real map of block indices for nodes
1972   for (int j = 0; j < _block.length(); j++) {
1973     Node* n = _block.at(j);
1974     set_bb_idx(n, j);
1975   }
1976 
1977   // Ensure extra info is allocated.
1978   initialize_bb();
1979 
1980 #ifndef PRODUCT
1981   if (TraceSuperWord) {
1982     print_bb();
1983     tty->print_cr("\ndata entry nodes: %s", _data_entry.length() > 0 ? "" : "NONE");
1984     for (int m = 0; m < _data_entry.length(); m++) {
1985       tty->print("%3d ", m);
1986       _data_entry.at(m)->dump();
1987     }
1988     tty->print_cr("\nmemory slices: %s", _mem_slice_head.length() > 0 ? "" : "NONE");
1989     for (int m = 0; m < _mem_slice_head.length(); m++) {
1990       tty->print("%3d ", m); _mem_slice_head.at(m)->dump();
1991       tty->print("    ");    _mem_slice_tail.at(m)->dump();
1992     }
1993   }
1994 #endif
1995   assert(rpo_idx == -1 && bb_ct == _block.length(), "all block members found");
1996   return (_mem_slice_head.length() > 0) || (reduction_uses > 0) || (_data_entry.length() > 0);
1997 }
1998 
1999 //------------------------------initialize_bb---------------------------
2000 // Initialize per node info
2001 void SuperWord::initialize_bb() {
2002   Node* last = _block.at(_block.length() - 1);
2003   grow_node_info(bb_idx(last));
2004 }
2005 
2006 //------------------------------bb_insert_after---------------------------
2007 // Insert n into block after pos
2008 void SuperWord::bb_insert_after(Node* n, int pos) {
2009   int n_pos = pos + 1;
2010   // Make room
2011   for (int i = _block.length() - 1; i >= n_pos; i--) {
2012     _block.at_put_grow(i+1, _block.at(i));
2013   }
2014   for (int j = _node_info.length() - 1; j >= n_pos; j--) {
2015     _node_info.at_put_grow(j+1, _node_info.at(j));
2016   }
2017   // Set value
2018   _block.at_put_grow(n_pos, n);
2019   _node_info.at_put_grow(n_pos, SWNodeInfo::initial);
2020   // Adjust map from node->_idx to _block index
2021   for (int i = n_pos; i < _block.length(); i++) {
2022     set_bb_idx(_block.at(i), i);
2023   }
2024 }
2025 
2026 //------------------------------compute_max_depth---------------------------
2027 // Compute max depth for expressions from beginning of block
2028 // Use to prune search paths during test for independence.
2029 void SuperWord::compute_max_depth() {
2030   int ct = 0;
2031   bool again;
2032   do {
2033     again = false;
2034     for (int i = 0; i < _block.length(); i++) {
2035       Node* n = _block.at(i);
2036       if (!n->is_Phi()) {
2037         int d_orig = depth(n);
2038         int d_in   = 0;
2039         for (DepPreds preds(n, _dg); !preds.done(); preds.next()) {
2040           Node* pred = preds.current();
2041           if (in_bb(pred)) {
2042             d_in = MAX2(d_in, depth(pred));
2043           }
2044         }
2045         if (d_in + 1 != d_orig) {
2046           set_depth(n, d_in + 1);
2047           again = true;
2048         }
2049       }
2050     }
2051     ct++;
2052   } while (again);
2053 #ifndef PRODUCT
2054   if (TraceSuperWord && Verbose)
2055     tty->print_cr("compute_max_depth iterated: %d times", ct);
2056 #endif
2057 }
2058 
2059 //-------------------------compute_vector_element_type-----------------------
2060 // Compute necessary vector element type for expressions
2061 // This propagates backwards a narrower integer type when the
2062 // upper bits of the value are not needed.
2063 // Example:  char a,b,c;  a = b + c;
2064 // Normally the type of the add is integer, but for packed character
2065 // operations the type of the add needs to be char.
2066 void SuperWord::compute_vector_element_type() {
2067 #ifndef PRODUCT
2068   if (TraceSuperWord && Verbose)
2069     tty->print_cr("\ncompute_velt_type:");
2070 #endif
2071 
2072   // Initial type
2073   for (int i = 0; i < _block.length(); i++) {
2074     Node* n = _block.at(i);
2075     set_velt_type(n, container_type(n));
2076   }
2077 
2078   // Propagate integer narrowed type backwards through operations
2079   // that don't depend on higher order bits
2080   for (int i = _block.length() - 1; i >= 0; i--) {
2081     Node* n = _block.at(i);
2082     // Only integer types need be examined
2083     const Type* vtn = velt_type(n);
2084     if (vtn->basic_type() == T_INT) {
2085       uint start, end;
2086       VectorNode::vector_operands(n, &start, &end);
2087 
2088       for (uint j = start; j < end; j++) {
2089         Node* in  = n->in(j);
2090         // Don't propagate through a memory
2091         if (!in->is_Mem() && in_bb(in) && velt_type(in)->basic_type() == T_INT &&
2092             data_size(n) < data_size(in)) {
2093           bool same_type = true;
2094           for (DUIterator_Fast kmax, k = in->fast_outs(kmax); k < kmax; k++) {
2095             Node *use = in->fast_out(k);
2096             if (!in_bb(use) || !same_velt_type(use, n)) {
2097               same_type = false;
2098               break;
2099             }
2100           }
2101           if (same_type) {
2102             // For right shifts of small integer types (bool, byte, char, short)
2103             // we need precise information about sign-ness. Only Load nodes have
2104             // this information because Store nodes are the same for signed and
2105             // unsigned values. And any arithmetic operation after a load may
2106             // expand a value to signed Int so such right shifts can't be used
2107             // because vector elements do not have upper bits of Int.
2108             const Type* vt = vtn;
2109             if (VectorNode::is_shift(in)) {
2110               Node* load = in->in(1);
2111               if (load->is_Load() && in_bb(load) && (velt_type(load)->basic_type() == T_INT)) {
2112                 vt = velt_type(load);
2113               } else if (in->Opcode() != Op_LShiftI) {
2114                 // Widen type to Int to avoid creation of right shift vector
2115                 // (align + data_size(s1) check in stmts_can_pack() will fail).
2116                 // Note, left shifts work regardless type.
2117                 vt = TypeInt::INT;
2118               }
2119             }
2120             set_velt_type(in, vt);
2121           }
2122         }
2123       }
2124     }
2125   }
2126 #ifndef PRODUCT
2127   if (TraceSuperWord && Verbose) {
2128     for (int i = 0; i < _block.length(); i++) {
2129       Node* n = _block.at(i);
2130       velt_type(n)->dump();
2131       tty->print("\t");
2132       n->dump();
2133     }
2134   }
2135 #endif
2136 }
2137 
2138 //------------------------------memory_alignment---------------------------
2139 // Alignment within a vector memory reference
2140 int SuperWord::memory_alignment(MemNode* s, int iv_adjust) {
2141   SWPointer p(s, this);
2142   if (!p.valid()) {
2143     return bottom_align;
2144   }
2145   int vw = vector_width_in_bytes(s);
2146   if (vw < 2) {
2147     return bottom_align; // No vectors for this type
2148   }
2149   int offset  = p.offset_in_bytes();
2150   offset     += iv_adjust*p.memory_size();
2151   int off_rem = offset % vw;
2152   int off_mod = off_rem >= 0 ? off_rem : off_rem + vw;
2153   return off_mod;
2154 }
2155 
2156 //---------------------------container_type---------------------------
2157 // Smallest type containing range of values
2158 const Type* SuperWord::container_type(Node* n) {
2159   if (n->is_Mem()) {
2160     BasicType bt = n->as_Mem()->memory_type();
2161     if (n->is_Store() && (bt == T_CHAR)) {
2162       // Use T_SHORT type instead of T_CHAR for stored values because any
2163       // preceding arithmetic operation extends values to signed Int.
2164       bt = T_SHORT;
2165     }
2166     if (n->Opcode() == Op_LoadUB) {
2167       // Adjust type for unsigned byte loads, it is important for right shifts.
2168       // T_BOOLEAN is used because there is no basic type representing type
2169       // TypeInt::UBYTE. Use of T_BOOLEAN for vectors is fine because only
2170       // size (one byte) and sign is important.
2171       bt = T_BOOLEAN;
2172     }
2173     return Type::get_const_basic_type(bt);
2174   }
2175   const Type* t = _igvn.type(n);
2176   if (t->basic_type() == T_INT) {
2177     // A narrow type of arithmetic operations will be determined by
2178     // propagating the type of memory operations.
2179     return TypeInt::INT;
2180   }
2181   return t;
2182 }
2183 
2184 bool SuperWord::same_velt_type(Node* n1, Node* n2) {
2185   const Type* vt1 = velt_type(n1);
2186   const Type* vt2 = velt_type(n2);
2187   if (vt1->basic_type() == T_INT && vt2->basic_type() == T_INT) {
2188     // Compare vectors element sizes for integer types.
2189     return data_size(n1) == data_size(n2);
2190   }
2191   return vt1 == vt2;
2192 }
2193 
2194 //------------------------------in_packset---------------------------
2195 // Are s1 and s2 in a pack pair and ordered as s1,s2?
2196 bool SuperWord::in_packset(Node* s1, Node* s2) {
2197   for (int i = 0; i < _packset.length(); i++) {
2198     Node_List* p = _packset.at(i);
2199     assert(p->size() == 2, "must be");
2200     if (p->at(0) == s1 && p->at(p->size()-1) == s2) {
2201       return true;
2202     }
2203   }
2204   return false;
2205 }
2206 
2207 //------------------------------in_pack---------------------------
2208 // Is s in pack p?
2209 Node_List* SuperWord::in_pack(Node* s, Node_List* p) {
2210   for (uint i = 0; i < p->size(); i++) {
2211     if (p->at(i) == s) {
2212       return p;
2213     }
2214   }
2215   return NULL;
2216 }
2217 
2218 //------------------------------remove_pack_at---------------------------
2219 // Remove the pack at position pos in the packset
2220 void SuperWord::remove_pack_at(int pos) {
2221   Node_List* p = _packset.at(pos);
2222   for (uint i = 0; i < p->size(); i++) {
2223     Node* s = p->at(i);
2224     set_my_pack(s, NULL);
2225   }
2226   _packset.remove_at(pos);
2227 }
2228 
2229 void SuperWord::packset_sort(int n) {
2230   // simple bubble sort so that we capitalize with O(n) when its already sorted
2231   while (n != 0) {
2232     bool swapped = false;
2233     for (int i = 1; i < n; i++) {
2234       Node_List* q_low = _packset.at(i-1);
2235       Node_List* q_i = _packset.at(i);
2236 
2237       // only swap when we find something to swap
2238       if (alignment(q_low->at(0)) > alignment(q_i->at(0))) {
2239         Node_List* t = q_i;
2240         *(_packset.adr_at(i)) = q_low;
2241         *(_packset.adr_at(i-1)) = q_i;
2242         swapped = true;
2243       }
2244     }
2245     if (swapped == false) break;
2246     n--;
2247   }
2248 }
2249 
2250 //------------------------------executed_first---------------------------
2251 // Return the node executed first in pack p.  Uses the RPO block list
2252 // to determine order.
2253 Node* SuperWord::executed_first(Node_List* p) {
2254   Node* n = p->at(0);
2255   int n_rpo = bb_idx(n);
2256   for (uint i = 1; i < p->size(); i++) {
2257     Node* s = p->at(i);
2258     int s_rpo = bb_idx(s);
2259     if (s_rpo < n_rpo) {
2260       n = s;
2261       n_rpo = s_rpo;
2262     }
2263   }
2264   return n;
2265 }
2266 
2267 //------------------------------executed_last---------------------------
2268 // Return the node executed last in pack p.
2269 Node* SuperWord::executed_last(Node_List* p) {
2270   Node* n = p->at(0);
2271   int n_rpo = bb_idx(n);
2272   for (uint i = 1; i < p->size(); i++) {
2273     Node* s = p->at(i);
2274     int s_rpo = bb_idx(s);
2275     if (s_rpo > n_rpo) {
2276       n = s;
2277       n_rpo = s_rpo;
2278     }
2279   }
2280   return n;
2281 }
2282 
2283 LoadNode::ControlDependency SuperWord::control_dependency(Node_List* p) {
2284   LoadNode::ControlDependency dep = LoadNode::DependsOnlyOnTest;
2285   for (uint i = 0; i < p->size(); i++) {
2286     Node* n = p->at(i);
2287     assert(n->is_Load(), "only meaningful for loads");
2288     if (!n->depends_only_on_test()) {
2289       dep = LoadNode::Pinned;
2290     }
2291   }
2292   return dep;
2293 }
2294 
2295 
2296 //----------------------------align_initial_loop_index---------------------------
2297 // Adjust pre-loop limit so that in main loop, a load/store reference
2298 // to align_to_ref will be a position zero in the vector.
2299 //   (iv + k) mod vector_align == 0
2300 void SuperWord::align_initial_loop_index(MemNode* align_to_ref) {
2301   CountedLoopNode *main_head = lp()->as_CountedLoop();
2302   assert(main_head->is_main_loop(), "");
2303   CountedLoopEndNode* pre_end = get_pre_loop_end(main_head);
2304   assert(pre_end != NULL, "we must have a correct pre-loop");
2305   Node *pre_opaq1 = pre_end->limit();
2306   assert(pre_opaq1->Opcode() == Op_Opaque1, "");
2307   Opaque1Node *pre_opaq = (Opaque1Node*)pre_opaq1;
2308   Node *lim0 = pre_opaq->in(1);
2309 
2310   // Where we put new limit calculations
2311   Node *pre_ctrl = pre_end->loopnode()->in(LoopNode::EntryControl);
2312 
2313   // Ensure the original loop limit is available from the
2314   // pre-loop Opaque1 node.
2315   Node *orig_limit = pre_opaq->original_loop_limit();
2316   assert(orig_limit != NULL && _igvn.type(orig_limit) != Type::TOP, "");
2317 
2318   SWPointer align_to_ref_p(align_to_ref, this);
2319   assert(align_to_ref_p.valid(), "sanity");
2320 
2321   // Given:
2322   //     lim0 == original pre loop limit
2323   //     V == v_align (power of 2)
2324   //     invar == extra invariant piece of the address expression
2325   //     e == offset [ +/- invar ]
2326   //
2327   // When reassociating expressions involving '%' the basic rules are:
2328   //     (a - b) % k == 0   =>  a % k == b % k
2329   // and:
2330   //     (a + b) % k == 0   =>  a % k == (k - b) % k
2331   //
2332   // For stride > 0 && scale > 0,
2333   //   Derive the new pre-loop limit "lim" such that the two constraints:
2334   //     (1) lim = lim0 + N           (where N is some positive integer < V)
2335   //     (2) (e + lim) % V == 0
2336   //   are true.
2337   //
2338   //   Substituting (1) into (2),
2339   //     (e + lim0 + N) % V == 0
2340   //   solve for N:
2341   //     N = (V - (e + lim0)) % V
2342   //   substitute back into (1), so that new limit
2343   //     lim = lim0 + (V - (e + lim0)) % V
2344   //
2345   // For stride > 0 && scale < 0
2346   //   Constraints:
2347   //     lim = lim0 + N
2348   //     (e - lim) % V == 0
2349   //   Solving for lim:
2350   //     (e - lim0 - N) % V == 0
2351   //     N = (e - lim0) % V
2352   //     lim = lim0 + (e - lim0) % V
2353   //
2354   // For stride < 0 && scale > 0
2355   //   Constraints:
2356   //     lim = lim0 - N
2357   //     (e + lim) % V == 0
2358   //   Solving for lim:
2359   //     (e + lim0 - N) % V == 0
2360   //     N = (e + lim0) % V
2361   //     lim = lim0 - (e + lim0) % V
2362   //
2363   // For stride < 0 && scale < 0
2364   //   Constraints:
2365   //     lim = lim0 - N
2366   //     (e - lim) % V == 0
2367   //   Solving for lim:
2368   //     (e - lim0 + N) % V == 0
2369   //     N = (V - (e - lim0)) % V
2370   //     lim = lim0 - (V - (e - lim0)) % V
2371 
2372   int vw = vector_width_in_bytes(align_to_ref);
2373   int stride   = iv_stride();
2374   int scale    = align_to_ref_p.scale_in_bytes();
2375   int elt_size = align_to_ref_p.memory_size();
2376   int v_align  = vw / elt_size;
2377   assert(v_align > 1, "sanity");
2378   int offset   = align_to_ref_p.offset_in_bytes() / elt_size;
2379   Node *offsn  = _igvn.intcon(offset);
2380 
2381   Node *e = offsn;
2382   if (align_to_ref_p.invar() != NULL) {
2383     // incorporate any extra invariant piece producing (offset +/- invar) >>> log2(elt)
2384     Node* log2_elt = _igvn.intcon(exact_log2(elt_size));
2385     Node* aref     = new URShiftINode(align_to_ref_p.invar(), log2_elt);
2386     _igvn.register_new_node_with_optimizer(aref);
2387     _phase->set_ctrl(aref, pre_ctrl);
2388     if (align_to_ref_p.negate_invar()) {
2389       e = new SubINode(e, aref);
2390     } else {
2391       e = new AddINode(e, aref);
2392     }
2393     _igvn.register_new_node_with_optimizer(e);
2394     _phase->set_ctrl(e, pre_ctrl);
2395   }
2396   if (vw > ObjectAlignmentInBytes) {
2397     // incorporate base e +/- base && Mask >>> log2(elt)
2398     Node* xbase = new CastP2XNode(NULL, align_to_ref_p.base());
2399     _igvn.register_new_node_with_optimizer(xbase);
2400 #ifdef _LP64
2401     xbase  = new ConvL2INode(xbase);
2402     _igvn.register_new_node_with_optimizer(xbase);
2403 #endif
2404     Node* mask = _igvn.intcon(vw-1);
2405     Node* masked_xbase  = new AndINode(xbase, mask);
2406     _igvn.register_new_node_with_optimizer(masked_xbase);
2407     Node* log2_elt = _igvn.intcon(exact_log2(elt_size));
2408     Node* bref     = new URShiftINode(masked_xbase, log2_elt);
2409     _igvn.register_new_node_with_optimizer(bref);
2410     _phase->set_ctrl(bref, pre_ctrl);
2411     e = new AddINode(e, bref);
2412     _igvn.register_new_node_with_optimizer(e);
2413     _phase->set_ctrl(e, pre_ctrl);
2414   }
2415 
2416   // compute e +/- lim0
2417   if (scale < 0) {
2418     e = new SubINode(e, lim0);
2419   } else {
2420     e = new AddINode(e, lim0);
2421   }
2422   _igvn.register_new_node_with_optimizer(e);
2423   _phase->set_ctrl(e, pre_ctrl);
2424 
2425   if (stride * scale > 0) {
2426     // compute V - (e +/- lim0)
2427     Node* va  = _igvn.intcon(v_align);
2428     e = new SubINode(va, e);
2429     _igvn.register_new_node_with_optimizer(e);
2430     _phase->set_ctrl(e, pre_ctrl);
2431   }
2432   // compute N = (exp) % V
2433   Node* va_msk = _igvn.intcon(v_align - 1);
2434   Node* N = new AndINode(e, va_msk);
2435   _igvn.register_new_node_with_optimizer(N);
2436   _phase->set_ctrl(N, pre_ctrl);
2437 
2438   //   substitute back into (1), so that new limit
2439   //     lim = lim0 + N
2440   Node* lim;
2441   if (stride < 0) {
2442     lim = new SubINode(lim0, N);
2443   } else {
2444     lim = new AddINode(lim0, N);
2445   }
2446   _igvn.register_new_node_with_optimizer(lim);
2447   _phase->set_ctrl(lim, pre_ctrl);
2448   Node* constrained =
2449     (stride > 0) ? (Node*) new MinINode(lim, orig_limit)
2450                  : (Node*) new MaxINode(lim, orig_limit);
2451   _igvn.register_new_node_with_optimizer(constrained);
2452   _phase->set_ctrl(constrained, pre_ctrl);
2453   _igvn.hash_delete(pre_opaq);
2454   pre_opaq->set_req(1, constrained);
2455 }
2456 
2457 //----------------------------get_pre_loop_end---------------------------
2458 // Find pre loop end from main loop.  Returns null if none.
2459 CountedLoopEndNode* SuperWord::get_pre_loop_end(CountedLoopNode *cl) {
2460   Node *ctrl = cl->in(LoopNode::EntryControl);
2461   if (!ctrl->is_IfTrue() && !ctrl->is_IfFalse()) return NULL;
2462   Node *iffm = ctrl->in(0);
2463   if (!iffm->is_If()) return NULL;
2464   Node *p_f = iffm->in(0);
2465   if (!p_f->is_IfFalse()) return NULL;
2466   if (!p_f->in(0)->is_CountedLoopEnd()) return NULL;
2467   CountedLoopEndNode *pre_end = p_f->in(0)->as_CountedLoopEnd();
2468   CountedLoopNode* loop_node = pre_end->loopnode();
2469   if (loop_node == NULL || !loop_node->is_pre_loop()) return NULL;
2470   return pre_end;
2471 }
2472 
2473 
2474 //------------------------------init---------------------------
2475 void SuperWord::init() {
2476   _dg.init();
2477   _packset.clear();
2478   _disjoint_ptrs.clear();
2479   _block.clear();
2480   _data_entry.clear();
2481   _mem_slice_head.clear();
2482   _mem_slice_tail.clear();
2483   _iteration_first.clear();
2484   _iteration_last.clear();
2485   _node_info.clear();
2486   _align_to_ref = NULL;
2487   _lpt = NULL;
2488   _lp = NULL;
2489   _bb = NULL;
2490   _iv = NULL;
2491   _race_possible = 0;
2492   _num_work_vecs = 0;
2493   _num_reductions = 0;
2494 }
2495 
2496 //------------------------------restart---------------------------
2497 void SuperWord::restart() {
2498   _dg.init();
2499   _packset.clear();
2500   _disjoint_ptrs.clear();
2501   _block.clear();
2502   _data_entry.clear();
2503   _mem_slice_head.clear();
2504   _mem_slice_tail.clear();
2505   _node_info.clear();
2506 }
2507 
2508 //------------------------------print_packset---------------------------
2509 void SuperWord::print_packset() {
2510 #ifndef PRODUCT
2511   tty->print_cr("packset");
2512   for (int i = 0; i < _packset.length(); i++) {
2513     tty->print_cr("Pack: %d", i);
2514     Node_List* p = _packset.at(i);
2515     print_pack(p);
2516   }
2517 #endif
2518 }
2519 
2520 //------------------------------print_pack---------------------------
2521 void SuperWord::print_pack(Node_List* p) {
2522   for (uint i = 0; i < p->size(); i++) {
2523     print_stmt(p->at(i));
2524   }
2525 }
2526 
2527 //------------------------------print_bb---------------------------
2528 void SuperWord::print_bb() {
2529 #ifndef PRODUCT
2530   tty->print_cr("\nBlock");
2531   for (int i = 0; i < _block.length(); i++) {
2532     Node* n = _block.at(i);
2533     tty->print("%d ", i);
2534     if (n) {
2535       n->dump();
2536     }
2537   }
2538 #endif
2539 }
2540 
2541 //------------------------------print_stmt---------------------------
2542 void SuperWord::print_stmt(Node* s) {
2543 #ifndef PRODUCT
2544   tty->print(" align: %d \t", alignment(s));
2545   s->dump();
2546 #endif
2547 }
2548 
2549 //------------------------------blank---------------------------
2550 char* SuperWord::blank(uint depth) {
2551   static char blanks[101];
2552   assert(depth < 101, "too deep");
2553   for (uint i = 0; i < depth; i++) blanks[i] = ' ';
2554   blanks[depth] = '\0';
2555   return blanks;
2556 }
2557 
2558 
2559 //==============================SWPointer===========================
2560 
2561 //----------------------------SWPointer------------------------
2562 SWPointer::SWPointer(MemNode* mem, SuperWord* slp) :
2563   _mem(mem), _slp(slp),  _base(NULL),  _adr(NULL),
2564   _scale(0), _offset(0), _invar(NULL), _negate_invar(false) {
2565 
2566   Node* adr = mem->in(MemNode::Address);
2567   if (!adr->is_AddP()) {
2568     assert(!valid(), "too complex");
2569     return;
2570   }
2571   // Match AddP(base, AddP(ptr, k*iv [+ invariant]), constant)
2572   Node* base = adr->in(AddPNode::Base);
2573   // The base address should be loop invariant
2574   if (!invariant(base)) {
2575     assert(!valid(), "base address is loop variant");
2576     return;
2577   }
2578   //unsafe reference could not be aligned appropriately without runtime checking
2579   if (base == NULL || base->bottom_type() == Type::TOP) {
2580     assert(!valid(), "unsafe access");
2581     return;
2582   }
2583   for (int i = 0; i < 3; i++) {
2584     if (!scaled_iv_plus_offset(adr->in(AddPNode::Offset))) {
2585       assert(!valid(), "too complex");
2586       return;
2587     }
2588     adr = adr->in(AddPNode::Address);
2589     if (base == adr || !adr->is_AddP()) {
2590       break; // stop looking at addp's
2591     }
2592   }
2593   _base = base;
2594   _adr  = adr;
2595   assert(valid(), "Usable");
2596 }
2597 
2598 // Following is used to create a temporary object during
2599 // the pattern match of an address expression.
2600 SWPointer::SWPointer(SWPointer* p) :
2601   _mem(p->_mem), _slp(p->_slp),  _base(NULL),  _adr(NULL),
2602   _scale(0), _offset(0), _invar(NULL), _negate_invar(false) {}
2603 
2604 //------------------------scaled_iv_plus_offset--------------------
2605 // Match: k*iv + offset
2606 // where: k is a constant that maybe zero, and
2607 //        offset is (k2 [+/- invariant]) where k2 maybe zero and invariant is optional
2608 bool SWPointer::scaled_iv_plus_offset(Node* n) {
2609   if (scaled_iv(n)) {
2610     return true;
2611   }
2612   if (offset_plus_k(n)) {
2613     return true;
2614   }
2615   int opc = n->Opcode();
2616   if (opc == Op_AddI) {
2617     if (scaled_iv(n->in(1)) && offset_plus_k(n->in(2))) {
2618       return true;
2619     }
2620     if (scaled_iv(n->in(2)) && offset_plus_k(n->in(1))) {
2621       return true;
2622     }
2623   } else if (opc == Op_SubI) {
2624     if (scaled_iv(n->in(1)) && offset_plus_k(n->in(2), true)) {
2625       return true;
2626     }
2627     if (scaled_iv(n->in(2)) && offset_plus_k(n->in(1))) {
2628       _scale *= -1;
2629       return true;
2630     }
2631   }
2632   return false;
2633 }
2634 
2635 //----------------------------scaled_iv------------------------
2636 // Match: k*iv where k is a constant that's not zero
2637 bool SWPointer::scaled_iv(Node* n) {
2638   if (_scale != 0) {
2639     return false;  // already found a scale
2640   }
2641   if (n == iv()) {
2642     _scale = 1;
2643     return true;
2644   }
2645   int opc = n->Opcode();
2646   if (opc == Op_MulI) {
2647     if (n->in(1) == iv() && n->in(2)->is_Con()) {
2648       _scale = n->in(2)->get_int();
2649       return true;
2650     } else if (n->in(2) == iv() && n->in(1)->is_Con()) {
2651       _scale = n->in(1)->get_int();
2652       return true;
2653     }
2654   } else if (opc == Op_LShiftI) {
2655     if (n->in(1) == iv() && n->in(2)->is_Con()) {
2656       _scale = 1 << n->in(2)->get_int();
2657       return true;
2658     }
2659   } else if (opc == Op_ConvI2L) {
2660     if (scaled_iv_plus_offset(n->in(1))) {
2661       return true;
2662     }
2663   } else if (opc == Op_LShiftL) {
2664     if (!has_iv() && _invar == NULL) {
2665       // Need to preserve the current _offset value, so
2666       // create a temporary object for this expression subtree.
2667       // Hacky, so should re-engineer the address pattern match.
2668       SWPointer tmp(this);
2669       if (tmp.scaled_iv_plus_offset(n->in(1))) {
2670         if (tmp._invar == NULL) {
2671           int mult = 1 << n->in(2)->get_int();
2672           _scale   = tmp._scale  * mult;
2673           _offset += tmp._offset * mult;
2674           return true;
2675         }
2676       }
2677     }
2678   }
2679   return false;
2680 }
2681 
2682 //----------------------------offset_plus_k------------------------
2683 // Match: offset is (k [+/- invariant])
2684 // where k maybe zero and invariant is optional, but not both.
2685 bool SWPointer::offset_plus_k(Node* n, bool negate) {
2686   int opc = n->Opcode();
2687   if (opc == Op_ConI) {
2688     _offset += negate ? -(n->get_int()) : n->get_int();
2689     return true;
2690   } else if (opc == Op_ConL) {
2691     // Okay if value fits into an int
2692     const TypeLong* t = n->find_long_type();
2693     if (t->higher_equal(TypeLong::INT)) {
2694       jlong loff = n->get_long();
2695       jint  off  = (jint)loff;
2696       _offset += negate ? -off : loff;
2697       return true;
2698     }
2699     return false;
2700   }
2701   if (_invar != NULL) return false; // already have an invariant
2702   if (opc == Op_AddI) {
2703     if (n->in(2)->is_Con() && invariant(n->in(1))) {
2704       _negate_invar = negate;
2705       _invar = n->in(1);
2706       _offset += negate ? -(n->in(2)->get_int()) : n->in(2)->get_int();
2707       return true;
2708     } else if (n->in(1)->is_Con() && invariant(n->in(2))) {
2709       _offset += negate ? -(n->in(1)->get_int()) : n->in(1)->get_int();
2710       _negate_invar = negate;
2711       _invar = n->in(2);
2712       return true;
2713     }
2714   }
2715   if (opc == Op_SubI) {
2716     if (n->in(2)->is_Con() && invariant(n->in(1))) {
2717       _negate_invar = negate;
2718       _invar = n->in(1);
2719       _offset += !negate ? -(n->in(2)->get_int()) : n->in(2)->get_int();
2720       return true;
2721     } else if (n->in(1)->is_Con() && invariant(n->in(2))) {
2722       _offset += negate ? -(n->in(1)->get_int()) : n->in(1)->get_int();
2723       _negate_invar = !negate;
2724       _invar = n->in(2);
2725       return true;
2726     }
2727   }
2728   if (invariant(n)) {
2729     _negate_invar = negate;
2730     _invar = n;
2731     return true;
2732   }
2733   return false;
2734 }
2735 
2736 //----------------------------print------------------------
2737 void SWPointer::print() {
2738 #ifndef PRODUCT
2739   tty->print("base: %d  adr: %d  scale: %d  offset: %d  invar: %c%d\n",
2740              _base != NULL ? _base->_idx : 0,
2741              _adr  != NULL ? _adr->_idx  : 0,
2742              _scale, _offset,
2743              _negate_invar?'-':'+',
2744              _invar != NULL ? _invar->_idx : 0);
2745 #endif
2746 }
2747 
2748 // ========================= OrderedPair =====================
2749 
2750 const OrderedPair OrderedPair::initial;
2751 
2752 // ========================= SWNodeInfo =====================
2753 
2754 const SWNodeInfo SWNodeInfo::initial;
2755 
2756 
2757 // ============================ DepGraph ===========================
2758 
2759 //------------------------------make_node---------------------------
2760 // Make a new dependence graph node for an ideal node.
2761 DepMem* DepGraph::make_node(Node* node) {
2762   DepMem* m = new (_arena) DepMem(node);
2763   if (node != NULL) {
2764     assert(_map.at_grow(node->_idx) == NULL, "one init only");
2765     _map.at_put_grow(node->_idx, m);
2766   }
2767   return m;
2768 }
2769 
2770 //------------------------------make_edge---------------------------
2771 // Make a new dependence graph edge from dpred -> dsucc
2772 DepEdge* DepGraph::make_edge(DepMem* dpred, DepMem* dsucc) {
2773   DepEdge* e = new (_arena) DepEdge(dpred, dsucc, dsucc->in_head(), dpred->out_head());
2774   dpred->set_out_head(e);
2775   dsucc->set_in_head(e);
2776   return e;
2777 }
2778 
2779 // ========================== DepMem ========================
2780 
2781 //------------------------------in_cnt---------------------------
2782 int DepMem::in_cnt() {
2783   int ct = 0;
2784   for (DepEdge* e = _in_head; e != NULL; e = e->next_in()) ct++;
2785   return ct;
2786 }
2787 
2788 //------------------------------out_cnt---------------------------
2789 int DepMem::out_cnt() {
2790   int ct = 0;
2791   for (DepEdge* e = _out_head; e != NULL; e = e->next_out()) ct++;
2792   return ct;
2793 }
2794 
2795 //------------------------------print-----------------------------
2796 void DepMem::print() {
2797 #ifndef PRODUCT
2798   tty->print("  DepNode %d (", _node->_idx);
2799   for (DepEdge* p = _in_head; p != NULL; p = p->next_in()) {
2800     Node* pred = p->pred()->node();
2801     tty->print(" %d", pred != NULL ? pred->_idx : 0);
2802   }
2803   tty->print(") [");
2804   for (DepEdge* s = _out_head; s != NULL; s = s->next_out()) {
2805     Node* succ = s->succ()->node();
2806     tty->print(" %d", succ != NULL ? succ->_idx : 0);
2807   }
2808   tty->print_cr(" ]");
2809 #endif
2810 }
2811 
2812 // =========================== DepEdge =========================
2813 
2814 //------------------------------DepPreds---------------------------
2815 void DepEdge::print() {
2816 #ifndef PRODUCT
2817   tty->print_cr("DepEdge: %d [ %d ]", _pred->node()->_idx, _succ->node()->_idx);
2818 #endif
2819 }
2820 
2821 // =========================== DepPreds =========================
2822 // Iterator over predecessor edges in the dependence graph.
2823 
2824 //------------------------------DepPreds---------------------------
2825 DepPreds::DepPreds(Node* n, DepGraph& dg) {
2826   _n = n;
2827   _done = false;
2828   if (_n->is_Store() || _n->is_Load()) {
2829     _next_idx = MemNode::Address;
2830     _end_idx  = n->req();
2831     _dep_next = dg.dep(_n)->in_head();
2832   } else if (_n->is_Mem()) {
2833     _next_idx = 0;
2834     _end_idx  = 0;
2835     _dep_next = dg.dep(_n)->in_head();
2836   } else {
2837     _next_idx = 1;
2838     _end_idx  = _n->req();
2839     _dep_next = NULL;
2840   }
2841   next();
2842 }
2843 
2844 //------------------------------next---------------------------
2845 void DepPreds::next() {
2846   if (_dep_next != NULL) {
2847     _current  = _dep_next->pred()->node();
2848     _dep_next = _dep_next->next_in();
2849   } else if (_next_idx < _end_idx) {
2850     _current  = _n->in(_next_idx++);
2851   } else {
2852     _done = true;
2853   }
2854 }
2855 
2856 // =========================== DepSuccs =========================
2857 // Iterator over successor edges in the dependence graph.
2858 
2859 //------------------------------DepSuccs---------------------------
2860 DepSuccs::DepSuccs(Node* n, DepGraph& dg) {
2861   _n = n;
2862   _done = false;
2863   if (_n->is_Load()) {
2864     _next_idx = 0;
2865     _end_idx  = _n->outcnt();
2866     _dep_next = dg.dep(_n)->out_head();
2867   } else if (_n->is_Mem() || _n->is_Phi() && _n->bottom_type() == Type::MEMORY) {
2868     _next_idx = 0;
2869     _end_idx  = 0;
2870     _dep_next = dg.dep(_n)->out_head();
2871   } else {
2872     _next_idx = 0;
2873     _end_idx  = _n->outcnt();
2874     _dep_next = NULL;
2875   }
2876   next();
2877 }
2878 
2879 //-------------------------------next---------------------------
2880 void DepSuccs::next() {
2881   if (_dep_next != NULL) {
2882     _current  = _dep_next->succ()->node();
2883     _dep_next = _dep_next->next_out();
2884   } else if (_next_idx < _end_idx) {
2885     _current  = _n->raw_out(_next_idx++);
2886   } else {
2887     _done = true;
2888   }
2889 }
2890 
2891 //
2892 // --------------------------------- vectorization/simd -----------------------------------
2893 //
2894 Node*  SuperWord::find_phi_for_mem_dep(LoadNode* ld) {
2895   assert(in_bb(ld), "must be in block");
2896   if (_clone_map.gen(ld->_idx) == _ii_first) {
2897 #ifndef PRODUCT
2898     if (_vector_loop_debug) {
2899       tty->print_cr("SuperWord::find_phi_for_mem_dep _clone_map.gen(ld->_idx)=%d",
2900                     _clone_map.gen(ld->_idx));
2901     }
2902 #endif
2903     return NULL; //we think that any ld in the first gen being vectorizable
2904   }
2905 
2906   Node* mem = ld->in(MemNode::Memory);
2907   if (mem->outcnt() <= 1) {
2908     // we don't want to remove the only edge from mem node to load
2909 #ifndef PRODUCT
2910     if (_vector_loop_debug) {
2911       tty->print_cr("SuperWord::find_phi_for_mem_dep input node %d to load %d has no other outputs and edge mem->load cannot be removed",
2912                     mem->_idx, ld->_idx);
2913       ld->dump();
2914       mem->dump();
2915     }
2916 #endif
2917     return NULL;
2918   }
2919   if (!in_bb(mem) || _clone_map.gen(mem->_idx) == _clone_map.gen(ld->_idx)) {
2920 #ifndef PRODUCT
2921     if (_vector_loop_debug) {
2922       tty->print_cr("SuperWord::find_phi_for_mem_dep _clone_map.gen(mem->_idx)=%d",
2923                     _clone_map.gen(mem->_idx));
2924     }
2925 #endif
2926     return NULL; // does not depend on loop volatile node or depends on the same generation
2927   }
2928 
2929   //otherwise first node should depend on mem-phi
2930   Node* first = first_node(ld);
2931   assert(first->is_Load(), "must be Load");
2932   Node* phi = first->as_Load()->in(MemNode::Memory);
2933   if (!phi->is_Phi() || phi->bottom_type() != Type::MEMORY) {
2934 #ifndef PRODUCT
2935     if (_vector_loop_debug) {
2936       tty->print_cr("SuperWord::find_phi_for_mem_dep load is not vectorizable node, since it's `first` does not take input from mem phi");
2937       ld->dump();
2938       first->dump();
2939     }
2940 #endif
2941     return NULL;
2942   }
2943 
2944   Node* tail = 0;
2945   for (int m = 0; m < _mem_slice_head.length(); m++) {
2946     if (_mem_slice_head.at(m) == phi) {
2947       tail = _mem_slice_tail.at(m);
2948     }
2949   }
2950   if (tail == 0) { //test that found phi is in the list  _mem_slice_head
2951 #ifndef PRODUCT
2952     if (_vector_loop_debug) {
2953       tty->print_cr("SuperWord::find_phi_for_mem_dep load %d is not vectorizable node, its phi %d is not _mem_slice_head",
2954                     ld->_idx, phi->_idx);
2955       ld->dump();
2956       phi->dump();
2957     }
2958 #endif
2959     return NULL;
2960   }
2961 
2962   // now all conditions are met
2963   return phi;
2964 }
2965 
2966 Node* SuperWord::first_node(Node* nd) {
2967   for (int ii = 0; ii < _iteration_first.length(); ii++) {
2968     Node* nnn = _iteration_first.at(ii);
2969     if (_clone_map.idx(nnn->_idx) == _clone_map.idx(nd->_idx)) {
2970 #ifndef PRODUCT
2971       if (_vector_loop_debug) {
2972         tty->print_cr("SuperWord::first_node: %d is the first iteration node for %d (_clone_map.idx(nnn->_idx) = %d)",
2973                       nnn->_idx, nd->_idx, _clone_map.idx(nnn->_idx));
2974       }
2975 #endif
2976       return nnn;
2977     }
2978   }
2979 
2980 #ifndef PRODUCT
2981   if (_vector_loop_debug) {
2982     tty->print_cr("SuperWord::first_node: did not find first iteration node for %d (_clone_map.idx(nd->_idx)=%d)",
2983                   nd->_idx, _clone_map.idx(nd->_idx));
2984   }
2985 #endif
2986   return 0;
2987 }
2988 
2989 Node* SuperWord::last_node(Node* nd) {
2990   for (int ii = 0; ii < _iteration_last.length(); ii++) {
2991     Node* nnn = _iteration_last.at(ii);
2992     if (_clone_map.idx(nnn->_idx) == _clone_map.idx(nd->_idx)) {
2993 #ifndef PRODUCT
2994       if (_vector_loop_debug) {
2995         tty->print_cr("SuperWord::last_node _clone_map.idx(nnn->_idx)=%d, _clone_map.idx(nd->_idx)=%d",
2996                       _clone_map.idx(nnn->_idx), _clone_map.idx(nd->_idx));
2997       }
2998 #endif
2999       return nnn;
3000     }
3001   }
3002   return 0;
3003 }
3004 
3005 int SuperWord::mark_generations() {
3006   Node *ii_err = 0, *tail_err;
3007   for (int i = 0; i < _mem_slice_head.length(); i++) {
3008     Node* phi  = _mem_slice_head.at(i);
3009     assert(phi->is_Phi(), "must be phi");
3010 
3011     Node* tail = _mem_slice_tail.at(i);
3012     if (_ii_last == -1) {
3013       tail_err = tail;
3014       _ii_last = _clone_map.gen(tail->_idx);
3015     }
3016     else if (_ii_last != _clone_map.gen(tail->_idx)) {
3017 #ifndef PRODUCT
3018       if (TraceSuperWord && Verbose) {
3019         tty->print_cr("SuperWord::mark_generations _ii_last error - found different generations in two tail nodes ");
3020         tail->dump();
3021         tail_err->dump();
3022       }
3023 #endif
3024       return -1;
3025     }
3026 
3027     // find first iteration in the loop
3028     for (DUIterator_Fast imax, i = phi->fast_outs(imax); i < imax; i++) {
3029       Node* ii = phi->fast_out(i);
3030       if (in_bb(ii) && ii->is_Store()) { // we speculate that normally Stores of one and one only generation have deps from mem phi
3031         if (_ii_first == -1) {
3032           ii_err = ii;
3033           _ii_first = _clone_map.gen(ii->_idx);
3034         } else if (_ii_first != _clone_map.gen(ii->_idx)) {
3035 #ifndef PRODUCT
3036           if (TraceSuperWord && Verbose) {
3037             tty->print_cr("SuperWord::mark_generations _ii_first error - found different generations in two nodes ");
3038             ii->dump();
3039             ii_err->dump();
3040           }
3041 #endif
3042           return -1; // this phi has Stores from different generations of unroll and cannot be simd/vectorized
3043         }
3044       }
3045     }//for (DUIterator_Fast imax,
3046   }//for (int i...
3047 
3048   if (_ii_first == -1 || _ii_last == -1) {
3049 #ifndef PRODUCT
3050     if (TraceSuperWord && Verbose) {
3051       tty->print_cr("SuperWord::mark_generations unknown error, something vent wrong");
3052     }
3053 #endif
3054     return -1; // something vent wrong
3055   }
3056   // collect nodes in the first and last generations
3057   assert(_iteration_first.length() == 0, "_iteration_first must be empty");
3058   assert(_iteration_last.length() == 0, "_iteration_last must be empty");
3059   for (int j = 0; j < _block.length(); j++) {
3060     Node* n = _block.at(j);
3061     node_idx_t gen = _clone_map.gen(n->_idx);
3062     if ((signed)gen == _ii_first) {
3063       _iteration_first.push(n);
3064     } else if ((signed)gen == _ii_last) {
3065       _iteration_last.push(n);
3066     }
3067   }
3068 
3069   // building order of iterations
3070   assert(_ii_order.length() == 0, "should be empty");
3071   if (ii_err != 0) {
3072     assert(in_bb(ii_err) && ii_err->is_Store(), "should be Store in bb");
3073     Node* nd = ii_err;
3074     while(_clone_map.gen(nd->_idx) != _ii_last) {
3075       _ii_order.push(_clone_map.gen(nd->_idx));
3076       bool found = false;
3077       for (DUIterator_Fast imax, i = nd->fast_outs(imax); i < imax; i++) {
3078         Node* use = nd->fast_out(i);
3079         if (_clone_map.idx(use->_idx) == _clone_map.idx(nd->_idx) && use->as_Store()->in(MemNode::Memory) == nd) {
3080           found = true;
3081           nd = use;
3082           break;
3083         }
3084       }//for
3085 
3086       if (found == false) {
3087 #ifndef PRODUCT
3088         if (TraceSuperWord && Verbose) {
3089           tty->print_cr("SuperWord::mark_generations: Cannot build order of iterations - no dependent Store for %d", nd->_idx);
3090         }
3091 #endif
3092         _ii_order.clear();
3093         return -1;
3094       }
3095     } //while
3096     _ii_order.push(_clone_map.gen(nd->_idx));
3097   }
3098 
3099 #ifndef PRODUCT
3100   if (_vector_loop_debug) {
3101     tty->print_cr("SuperWord::mark_generations");
3102     tty->print_cr("First generation (%d) nodes:", _ii_first);
3103     for (int ii = 0; ii < _iteration_first.length(); ii++)  _iteration_first.at(ii)->dump();
3104     tty->print_cr("Last generation (%d) nodes:", _ii_last);
3105     for (int ii = 0; ii < _iteration_last.length(); ii++)  _iteration_last.at(ii)->dump();
3106     tty->print_cr(" ");
3107 
3108     tty->print("SuperWord::List of generations: ");
3109     for (int jj = 0; jj < _ii_order.length(); ++jj) {
3110       tty->print("%d:%d ", jj, _ii_order.at(jj));
3111     }
3112     tty->print_cr(" ");
3113   }
3114 #endif
3115 
3116   return _ii_first;
3117 }
3118 
3119 bool SuperWord::fix_commutative_inputs(Node* gold, Node* fix) {
3120   assert(gold->is_Add() && fix->is_Add() || gold->is_Mul() && fix->is_Mul(), "should be only Add or Mul nodes");
3121   assert(_clone_map.idx(gold->_idx) == _clone_map.idx(fix->_idx), "should be clones of the same node");
3122   Node* gin1 = gold->in(1);
3123   Node* gin2 = gold->in(2);
3124   Node* fin1 = fix->in(1);
3125   Node* fin2 = fix->in(2);
3126   bool swapped = false;
3127 
3128   if (in_bb(gin1) && in_bb(gin2) && in_bb(fin1) && in_bb(fin1)) {
3129     if (_clone_map.idx(gin1->_idx) == _clone_map.idx(fin1->_idx) &&
3130         _clone_map.idx(gin2->_idx) == _clone_map.idx(fin2->_idx)) {
3131       return true; // nothing to fix
3132     }
3133     if (_clone_map.idx(gin1->_idx) == _clone_map.idx(fin2->_idx) &&
3134         _clone_map.idx(gin2->_idx) == _clone_map.idx(fin1->_idx)) {
3135       fix->swap_edges(1, 2);
3136       swapped = true;
3137     }
3138   }
3139   // at least one input comes from outside of bb
3140   if (gin1->_idx == fin1->_idx)  {
3141     return true; // nothing to fix
3142   }
3143   if (!swapped && (gin1->_idx == fin2->_idx || gin2->_idx == fin1->_idx))  { //swapping is expensive, check condition first
3144     fix->swap_edges(1, 2);
3145     swapped = true;
3146   }
3147 
3148   if (swapped) {
3149 #ifndef PRODUCT
3150     if (_vector_loop_debug) {
3151       tty->print_cr("SuperWord::fix_commutative_inputs: fixed node %d", fix->_idx);
3152     }
3153 #endif
3154     return true;
3155   }
3156 
3157 #ifndef PRODUCT
3158   if (TraceSuperWord && Verbose) {
3159     tty->print_cr("SuperWord::fix_commutative_inputs: cannot fix node %d", fix->_idx);
3160   }
3161 #endif
3162   return false;
3163 }
3164 
3165 bool SuperWord::pack_parallel() {
3166 #ifndef PRODUCT
3167   if (_vector_loop_debug) {
3168     tty->print_cr("SuperWord::pack_parallel: START");
3169   }
3170 #endif
3171 
3172   _packset.clear();
3173 
3174   for (int ii = 0; ii < _iteration_first.length(); ii++) {
3175     Node* nd = _iteration_first.at(ii);
3176     if (in_bb(nd) && (nd->is_Load() || nd->is_Store() || nd->is_Add() || nd->is_Mul())) {
3177       Node_List* pk = new Node_List();
3178       pk->push(nd);
3179       for (int gen = 1; gen < _ii_order.length(); ++gen) {
3180         for (int kk = 0; kk < _block.length(); kk++) {
3181           Node* clone = _block.at(kk);
3182           if (_clone_map.idx(clone->_idx) == _clone_map.idx(nd->_idx) &&
3183               _clone_map.gen(clone->_idx) == _ii_order.at(gen)) {
3184             if (nd->is_Add() || nd->is_Mul()) {
3185               fix_commutative_inputs(nd, clone);
3186             }
3187             pk->push(clone);
3188             if (pk->size() == 4) {
3189               _packset.append(pk);
3190 #ifndef PRODUCT
3191               if (_vector_loop_debug) {
3192                 tty->print_cr("SuperWord::pack_parallel: added pack ");
3193                 pk->dump();
3194               }
3195 #endif
3196               if (_clone_map.gen(clone->_idx) != _ii_last) {
3197                 pk = new Node_List();
3198               }
3199             }
3200             break;
3201           }
3202         }
3203       }//for
3204     }//if
3205   }//for
3206 
3207 #ifndef PRODUCT
3208   if (_vector_loop_debug) {
3209     tty->print_cr("SuperWord::pack_parallel: END");
3210   }
3211 #endif
3212 
3213   return true;
3214 }
3215 
3216 bool SuperWord::hoist_loads_in_graph() {
3217   GrowableArray<Node*> loads;
3218 
3219 #ifndef PRODUCT
3220   if (_vector_loop_debug) {
3221     tty->print_cr("SuperWord::hoist_loads_in_graph: total number _mem_slice_head.length() = %d", _mem_slice_head.length());
3222   }
3223 #endif
3224 
3225   for (int i = 0; i < _mem_slice_head.length(); i++) {
3226     Node* n = _mem_slice_head.at(i);
3227     if ( !in_bb(n) || !n->is_Phi() || n->bottom_type() != Type::MEMORY) {
3228 #ifndef PRODUCT
3229       if (TraceSuperWord && Verbose) {
3230         tty->print_cr("SuperWord::hoist_loads_in_graph: skipping unexpected node n=%d", n->_idx);
3231       }
3232 #endif
3233       continue;
3234     }
3235 
3236 #ifndef PRODUCT
3237     if (_vector_loop_debug) {
3238       tty->print_cr("SuperWord::hoist_loads_in_graph: processing phi %d  = _mem_slice_head.at(%d);", n->_idx, i);
3239     }
3240 #endif
3241 
3242     for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3243       Node* ld = n->fast_out(i);
3244       if (ld->is_Load() && ld->as_Load()->in(MemNode::Memory) == n && in_bb(ld)) {
3245         for (int i = 0; i < _block.length(); i++) {
3246           Node* ld2 = _block.at(i);
3247           if (ld2->is_Load() &&
3248               _clone_map.idx(ld->_idx) == _clone_map.idx(ld2->_idx) &&
3249               _clone_map.gen(ld->_idx) != _clone_map.gen(ld2->_idx)) { // <= do not collect the first generation ld
3250 #ifndef PRODUCT
3251             if (_vector_loop_debug) {
3252               tty->print_cr("SuperWord::hoist_loads_in_graph: will try to hoist load ld2->_idx=%d, cloned from %d (ld->_idx=%d)",
3253                             ld2->_idx, _clone_map.idx(ld->_idx), ld->_idx);
3254             }
3255 #endif
3256             // could not do on-the-fly, since iterator is immutable
3257             loads.push(ld2);
3258           }
3259         }// for
3260       }//if
3261     }//for (DUIterator_Fast imax,
3262   }//for (int i = 0; i
3263 
3264   for (int i = 0; i < loads.length(); i++) {
3265     LoadNode* ld = loads.at(i)->as_Load();
3266     Node* phi = find_phi_for_mem_dep(ld);
3267     if (phi != NULL) {
3268 #ifndef PRODUCT
3269       if (_vector_loop_debug) {
3270         tty->print_cr("SuperWord::hoist_loads_in_graph replacing MemNode::Memory(%d) edge in %d with one from %d",
3271                       MemNode::Memory, ld->_idx, phi->_idx);
3272       }
3273 #endif
3274       _igvn.replace_input_of(ld, MemNode::Memory, phi);
3275     }
3276   }//for
3277 
3278   restart(); // invalidate all basic structures, since we rebuilt the graph
3279 
3280 #ifndef PRODUCT
3281   if (TraceSuperWord && Verbose) {
3282     tty->print_cr("\nSuperWord::hoist_loads_in_graph() the graph was rebuilt, all structures invalidated and need rebuild");
3283   }
3284 #endif
3285   return true;
3286 }
3287