1 /*
   2  * Copyright (c) 2007, 2013, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  */
  23 
  24 #include "precompiled.hpp"
  25 #include "compiler/compileLog.hpp"
  26 #include "libadt/vectset.hpp"
  27 #include "memory/allocation.inline.hpp"
  28 #include "opto/addnode.hpp"
  29 #include "opto/callnode.hpp"
  30 #include "opto/divnode.hpp"
  31 #include "opto/matcher.hpp"
  32 #include "opto/memnode.hpp"
  33 #include "opto/mulnode.hpp"
  34 #include "opto/opcodes.hpp"
  35 #include "opto/superword.hpp"
  36 #include "opto/vectornode.hpp"
  37 
  38 //
  39 //                  S U P E R W O R D   T R A N S F O R M
  40 //=============================================================================
  41 
  42 //------------------------------SuperWord---------------------------
  43 SuperWord::SuperWord(PhaseIdealLoop* phase) :
  44   _phase(phase),
  45   _igvn(phase->_igvn),
  46   _arena(phase->C->comp_arena()),
  47   _packset(arena(), 8,  0, NULL),         // packs for the current block
  48   _bb_idx(arena(), (int)(1.10 * phase->C->unique()), 0, 0), // node idx to index in bb
  49   _block(arena(), 8,  0, NULL),           // nodes in current block
  50   _data_entry(arena(), 8,  0, NULL),      // nodes with all inputs from outside
  51   _mem_slice_head(arena(), 8,  0, NULL),  // memory slice heads
  52   _mem_slice_tail(arena(), 8,  0, NULL),  // memory slice tails
  53   _node_info(arena(), 8,  0, SWNodeInfo::initial), // info needed per node
  54   _align_to_ref(NULL),                    // memory reference to align vectors to
  55   _disjoint_ptrs(arena(), 8,  0, OrderedPair::initial), // runtime disambiguated pointer pairs
  56   _dg(_arena),                            // dependence graph
  57   _visited(arena()),                      // visited node set
  58   _post_visited(arena()),                 // post visited node set
  59   _n_idx_list(arena(), 8),                // scratch list of (node,index) pairs
  60   _stk(arena(), 8, 0, NULL),              // scratch stack of nodes
  61   _nlist(arena(), 8, 0, NULL),            // scratch list of nodes
  62   _lpt(NULL),                             // loop tree node
  63   _lp(NULL),                              // LoopNode
  64   _bb(NULL),                              // basic block
  65   _iv(NULL)                               // induction var
  66 {}
  67 
  68 //------------------------------transform_loop---------------------------
  69 void SuperWord::transform_loop(IdealLoopTree* lpt) {
  70   assert(UseSuperWord, "should be");
  71   // Do vectors exist on this architecture?
  72   if (Matcher::vector_width_in_bytes(T_BYTE) < 2) return;
  73 
  74   assert(lpt->_head->is_CountedLoop(), "must be");
  75   CountedLoopNode *cl = lpt->_head->as_CountedLoop();
  76 
  77   if (!cl->is_valid_counted_loop()) return; // skip malformed counted loop
  78 
  79   if (!cl->is_main_loop() ) return; // skip normal, pre, and post loops
  80 
  81   // Check for no control flow in body (other than exit)
  82   Node *cl_exit = cl->loopexit();
  83   if (cl_exit->in(0) != lpt->_head) return;
  84 
  85   // Make sure the are no extra control users of the loop backedge
  86   if (cl->back_control()->outcnt() != 1) {
  87     return;
  88   }
  89 
  90   // Check for pre-loop ending with CountedLoopEnd(Bool(Cmp(x,Opaque1(limit))))
  91   CountedLoopEndNode* pre_end = get_pre_loop_end(cl);
  92   if (pre_end == NULL) return;
  93   Node *pre_opaq1 = pre_end->limit();
  94   if (pre_opaq1->Opcode() != Op_Opaque1) return;
  95 
  96   init(); // initialize data structures
  97 
  98   set_lpt(lpt);
  99   set_lp(cl);
 100 
 101   // For now, define one block which is the entire loop body
 102   set_bb(cl);
 103 
 104   assert(_packset.length() == 0, "packset must be empty");
 105   SLP_extract();
 106 }
 107 
 108 //------------------------------SLP_extract---------------------------
 109 // Extract the superword level parallelism
 110 //
 111 // 1) A reverse post-order of nodes in the block is constructed.  By scanning
 112 //    this list from first to last, all definitions are visited before their uses.
 113 //
 114 // 2) A point-to-point dependence graph is constructed between memory references.
 115 //    This simplies the upcoming "independence" checker.
 116 //
 117 // 3) The maximum depth in the node graph from the beginning of the block
 118 //    to each node is computed.  This is used to prune the graph search
 119 //    in the independence checker.
 120 //
 121 // 4) For integer types, the necessary bit width is propagated backwards
 122 //    from stores to allow packed operations on byte, char, and short
 123 //    integers.  This reverses the promotion to type "int" that javac
 124 //    did for operations like: char c1,c2,c3;  c1 = c2 + c3.
 125 //
 126 // 5) One of the memory references is picked to be an aligned vector reference.
 127 //    The pre-loop trip count is adjusted to align this reference in the
 128 //    unrolled body.
 129 //
 130 // 6) The initial set of pack pairs is seeded with memory references.
 131 //
 132 // 7) The set of pack pairs is extended by following use->def and def->use links.
 133 //
 134 // 8) The pairs are combined into vector sized packs.
 135 //
 136 // 9) Reorder the memory slices to co-locate members of the memory packs.
 137 //
 138 // 10) Generate ideal vector nodes for the final set of packs and where necessary,
 139 //    inserting scalar promotion, vector creation from multiple scalars, and
 140 //    extraction of scalar values from vectors.
 141 //
 142 void SuperWord::SLP_extract() {
 143 
 144   // Ready the block
 145 
 146   if (!construct_bb())
 147     return; // Exit if no interesting nodes or complex graph.
 148 
 149   dependence_graph();
 150 
 151   compute_max_depth();
 152 
 153   compute_vector_element_type();
 154 
 155   // Attempt vectorization
 156 
 157   find_adjacent_refs();
 158 
 159   extend_packlist();
 160 
 161   combine_packs();
 162 
 163   construct_my_pack_map();
 164 
 165   filter_packs();
 166 
 167   schedule();
 168 
 169   output();
 170 }
 171 
 172 //------------------------------find_adjacent_refs---------------------------
 173 // Find the adjacent memory references and create pack pairs for them.
 174 // This is the initial set of packs that will then be extended by
 175 // following use->def and def->use links.  The align positions are
 176 // assigned relative to the reference "align_to_ref"
 177 void SuperWord::find_adjacent_refs() {
 178   // Get list of memory operations
 179   Node_List memops;
 180   for (int i = 0; i < _block.length(); i++) {
 181     Node* n = _block.at(i);
 182     if (n->is_Mem() && !n->is_LoadStore() && in_bb(n) &&
 183         is_java_primitive(n->as_Mem()->memory_type())) {
 184       int align = memory_alignment(n->as_Mem(), 0);
 185       if (align != bottom_align) {
 186         memops.push(n);
 187       }
 188     }
 189   }
 190 
 191   Node_List align_to_refs;
 192   int best_iv_adjustment = 0;
 193   MemNode* best_align_to_mem_ref = NULL;
 194 
 195   while (memops.size() != 0) {
 196     // Find a memory reference to align to.
 197     MemNode* mem_ref = find_align_to_ref(memops);
 198     if (mem_ref == NULL) break;
 199     align_to_refs.push(mem_ref);
 200     int iv_adjustment = get_iv_adjustment(mem_ref);
 201 
 202     if (best_align_to_mem_ref == NULL) {
 203       // Set memory reference which is the best from all memory operations
 204       // to be used for alignment. The pre-loop trip count is modified to align
 205       // this reference to a vector-aligned address.
 206       best_align_to_mem_ref = mem_ref;
 207       best_iv_adjustment = iv_adjustment;
 208     }
 209 
 210     SWPointer align_to_ref_p(mem_ref, this);
 211     // Set alignment relative to "align_to_ref" for all related memory operations.
 212     for (int i = memops.size() - 1; i >= 0; i--) {
 213       MemNode* s = memops.at(i)->as_Mem();
 214       if (isomorphic(s, mem_ref)) {
 215         SWPointer p2(s, this);
 216         if (p2.comparable(align_to_ref_p)) {
 217           int align = memory_alignment(s, iv_adjustment);
 218           set_alignment(s, align);
 219         }
 220       }
 221     }
 222 
 223     // Create initial pack pairs of memory operations for which
 224     // alignment is set and vectors will be aligned.
 225     bool create_pack = true;
 226     if (memory_alignment(mem_ref, best_iv_adjustment) == 0) {
 227       if (!Matcher::misaligned_vectors_ok()) {
 228         int vw = vector_width(mem_ref);
 229         int vw_best = vector_width(best_align_to_mem_ref);
 230         if (vw > vw_best) {
 231           // Do not vectorize a memory access with more elements per vector
 232           // if unaligned memory access is not allowed because number of
 233           // iterations in pre-loop will be not enough to align it.
 234           create_pack = false;
 235         } else {
 236           SWPointer p2(best_align_to_mem_ref, this);
 237           if (align_to_ref_p.invar() != p2.invar()) {
 238             // Do not vectorize memory accesses with different invariants
 239             // if unaligned memory accesses are not allowed.
 240             create_pack = false;
 241           }
 242         }
 243       }
 244     } else {
 245       if (same_velt_type(mem_ref, best_align_to_mem_ref)) {
 246         // Can't allow vectorization of unaligned memory accesses with the
 247         // same type since it could be overlapped accesses to the same array.
 248         create_pack = false;
 249       } else {
 250         // Allow independent (different type) unaligned memory operations
 251         // if HW supports them.
 252         if (!Matcher::misaligned_vectors_ok()) {
 253           create_pack = false;
 254         } else {
 255           // Check if packs of the same memory type but
 256           // with a different alignment were created before.
 257           for (uint i = 0; i < align_to_refs.size(); i++) {
 258             MemNode* mr = align_to_refs.at(i)->as_Mem();
 259             if (same_velt_type(mr, mem_ref) &&
 260                 memory_alignment(mr, iv_adjustment) != 0)
 261               create_pack = false;
 262           }
 263         }
 264       }
 265     }
 266     if (create_pack) {
 267       for (uint i = 0; i < memops.size(); i++) {
 268         Node* s1 = memops.at(i);
 269         int align = alignment(s1);
 270         if (align == top_align) continue;
 271         for (uint j = 0; j < memops.size(); j++) {
 272           Node* s2 = memops.at(j);
 273           if (alignment(s2) == top_align) continue;
 274           if (s1 != s2 && are_adjacent_refs(s1, s2)) {
 275             if (stmts_can_pack(s1, s2, align)) {
 276               Node_List* pair = new Node_List();
 277               pair->push(s1);
 278               pair->push(s2);
 279               _packset.append(pair);
 280             }
 281           }
 282         }
 283       }
 284     } else { // Don't create unaligned pack
 285       // First, remove remaining memory ops of the same type from the list.
 286       for (int i = memops.size() - 1; i >= 0; i--) {
 287         MemNode* s = memops.at(i)->as_Mem();
 288         if (same_velt_type(s, mem_ref)) {
 289           memops.remove(i);
 290         }
 291       }
 292 
 293       // Second, remove already constructed packs of the same type.
 294       for (int i = _packset.length() - 1; i >= 0; i--) {
 295         Node_List* p = _packset.at(i);
 296         MemNode* s = p->at(0)->as_Mem();
 297         if (same_velt_type(s, mem_ref)) {
 298           remove_pack_at(i);
 299         }
 300       }
 301 
 302       // If needed find the best memory reference for loop alignment again.
 303       if (same_velt_type(mem_ref, best_align_to_mem_ref)) {
 304         // Put memory ops from remaining packs back on memops list for
 305         // the best alignment search.
 306         uint orig_msize = memops.size();
 307         for (int i = 0; i < _packset.length(); i++) {
 308           Node_List* p = _packset.at(i);
 309           MemNode* s = p->at(0)->as_Mem();
 310           assert(!same_velt_type(s, mem_ref), "sanity");
 311           memops.push(s);
 312         }
 313         MemNode* best_align_to_mem_ref = find_align_to_ref(memops);
 314         if (best_align_to_mem_ref == NULL) break;
 315         best_iv_adjustment = get_iv_adjustment(best_align_to_mem_ref);
 316         // Restore list.
 317         while (memops.size() > orig_msize)
 318           (void)memops.pop();
 319       }
 320     } // unaligned memory accesses
 321 
 322     // Remove used mem nodes.
 323     for (int i = memops.size() - 1; i >= 0; i--) {
 324       MemNode* m = memops.at(i)->as_Mem();
 325       if (alignment(m) != top_align) {
 326         memops.remove(i);
 327       }
 328     }
 329 
 330   } // while (memops.size() != 0
 331   set_align_to_ref(best_align_to_mem_ref);
 332 
 333 #ifndef PRODUCT
 334   if (TraceSuperWord) {
 335     tty->print_cr("\nAfter find_adjacent_refs");
 336     print_packset();
 337   }
 338 #endif
 339 }
 340 
 341 //------------------------------find_align_to_ref---------------------------
 342 // Find a memory reference to align the loop induction variable to.
 343 // Looks first at stores then at loads, looking for a memory reference
 344 // with the largest number of references similar to it.
 345 MemNode* SuperWord::find_align_to_ref(Node_List &memops) {
 346   GrowableArray<int> cmp_ct(arena(), memops.size(), memops.size(), 0);
 347 
 348   // Count number of comparable memory ops
 349   for (uint i = 0; i < memops.size(); i++) {
 350     MemNode* s1 = memops.at(i)->as_Mem();
 351     SWPointer p1(s1, this);
 352     // Discard if pre loop can't align this reference
 353     if (!ref_is_alignable(p1)) {
 354       *cmp_ct.adr_at(i) = 0;
 355       continue;
 356     }
 357     for (uint j = i+1; j < memops.size(); j++) {
 358       MemNode* s2 = memops.at(j)->as_Mem();
 359       if (isomorphic(s1, s2)) {
 360         SWPointer p2(s2, this);
 361         if (p1.comparable(p2)) {
 362           (*cmp_ct.adr_at(i))++;
 363           (*cmp_ct.adr_at(j))++;
 364         }
 365       }
 366     }
 367   }
 368 
 369   // Find Store (or Load) with the greatest number of "comparable" references,
 370   // biggest vector size, smallest data size and smallest iv offset.
 371   int max_ct        = 0;
 372   int max_vw        = 0;
 373   int max_idx       = -1;
 374   int min_size      = max_jint;
 375   int min_iv_offset = max_jint;
 376   for (uint j = 0; j < memops.size(); j++) {
 377     MemNode* s = memops.at(j)->as_Mem();
 378     if (s->is_Store()) {
 379       int vw = vector_width_in_bytes(s);
 380       assert(vw > 1, "sanity");
 381       SWPointer p(s, this);
 382       if (cmp_ct.at(j) >  max_ct ||
 383           cmp_ct.at(j) == max_ct &&
 384             (vw >  max_vw ||
 385              vw == max_vw &&
 386               (data_size(s) <  min_size ||
 387                data_size(s) == min_size &&
 388                  (p.offset_in_bytes() < min_iv_offset)))) {
 389         max_ct = cmp_ct.at(j);
 390         max_vw = vw;
 391         max_idx = j;
 392         min_size = data_size(s);
 393         min_iv_offset = p.offset_in_bytes();
 394       }
 395     }
 396   }
 397   // If no stores, look at loads
 398   if (max_ct == 0) {
 399     for (uint j = 0; j < memops.size(); j++) {
 400       MemNode* s = memops.at(j)->as_Mem();
 401       if (s->is_Load()) {
 402         int vw = vector_width_in_bytes(s);
 403         assert(vw > 1, "sanity");
 404         SWPointer p(s, this);
 405         if (cmp_ct.at(j) >  max_ct ||
 406             cmp_ct.at(j) == max_ct &&
 407               (vw >  max_vw ||
 408                vw == max_vw &&
 409                 (data_size(s) <  min_size ||
 410                  data_size(s) == min_size &&
 411                    (p.offset_in_bytes() < min_iv_offset)))) {
 412           max_ct = cmp_ct.at(j);
 413           max_vw = vw;
 414           max_idx = j;
 415           min_size = data_size(s);
 416           min_iv_offset = p.offset_in_bytes();
 417         }
 418       }
 419     }
 420   }
 421 
 422 #ifdef ASSERT
 423   if (TraceSuperWord && Verbose) {
 424     tty->print_cr("\nVector memops after find_align_to_refs");
 425     for (uint i = 0; i < memops.size(); i++) {
 426       MemNode* s = memops.at(i)->as_Mem();
 427       s->dump();
 428     }
 429   }
 430 #endif
 431 
 432   if (max_ct > 0) {
 433 #ifdef ASSERT
 434     if (TraceSuperWord) {
 435       tty->print("\nVector align to node: ");
 436       memops.at(max_idx)->as_Mem()->dump();
 437     }
 438 #endif
 439     return memops.at(max_idx)->as_Mem();
 440   }
 441   return NULL;
 442 }
 443 
 444 //------------------------------ref_is_alignable---------------------------
 445 // Can the preloop align the reference to position zero in the vector?
 446 bool SuperWord::ref_is_alignable(SWPointer& p) {
 447   if (!p.has_iv()) {
 448     return true;   // no induction variable
 449   }
 450   CountedLoopEndNode* pre_end = get_pre_loop_end(lp()->as_CountedLoop());
 451   assert(pre_end != NULL, "we must have a correct pre-loop");
 452   assert(pre_end->stride_is_con(), "pre loop stride is constant");
 453   int preloop_stride = pre_end->stride_con();
 454 
 455   int span = preloop_stride * p.scale_in_bytes();
 456   int mem_size = p.memory_size();
 457   int offset   = p.offset_in_bytes();
 458   // Stride one accesses are alignable if offset is aligned to memory operation size.
 459   // Offset can be unaligned when UseUnalignedAccesses is used.
 460   if (ABS(span) == mem_size && (ABS(offset) % mem_size) == 0) {
 461     return true;
 462   }
 463   // If the initial offset from start of the object is computable,
 464   // check if the pre-loop can align the final offset accordingly.
 465   //
 466   // In other words: Can we find an i such that the offset
 467   // after i pre-loop iterations is aligned to vw?
 468   //   (init_offset + pre_loop) % vw == 0              (1)
 469   // where
 470   //   pre_loop = i * span
 471   // is the number of bytes added to the offset by i pre-loop iterations.
 472   //
 473   // For this to hold we need pre_loop to increase init_offset by
 474   //   pre_loop = vw - (init_offset % vw)
 475   //
 476   // This is only possible if pre_loop is divisible by span because each
 477   // pre-loop iteration increases the initial offset by 'span' bytes:
 478   //   (vw - (init_offset % vw)) % span == 0
 479   //
 480   int vw = vector_width_in_bytes(p.mem());
 481   assert(vw > 1, "sanity");
 482   Node* init_nd = pre_end->init_trip();
 483   if (init_nd->is_Con() && p.invar() == NULL) {
 484     int init = init_nd->bottom_type()->is_int()->get_con();
 485     int init_offset = init * p.scale_in_bytes() + offset;
 486     if (init_offset < 0) { // negative offset from object start?
 487       return false;        // may happen in dead loop
 488     }
 489     if (vw % span == 0) {
 490       // If vm is a multiple of span, we use formula (1).
 491       if (span > 0) {
 492         return (vw - (init_offset % vw)) % span == 0;
 493       } else {
 494         assert(span < 0, "nonzero stride * scale");
 495         return (init_offset % vw) % -span == 0;
 496       }
 497     } else if (span % vw == 0) {
 498       // If span is a multiple of vw, we can simplify formula (1) to:
 499       //   (init_offset + i * span) % vw == 0
 500       //     =>
 501       //   (init_offset % vw) + ((i * span) % vw) == 0
 502       //     =>
 503       //   init_offset % vw == 0
 504       //
 505       // Because we add a multiple of vw to the initial offset, the final
 506       // offset is a multiple of vw if and only if init_offset is a multiple.
 507       //
 508       return (init_offset % vw) == 0;
 509     }
 510   }
 511   return false;
 512 }
 513 
 514 //---------------------------get_iv_adjustment---------------------------
 515 // Calculate loop's iv adjustment for this memory ops.
 516 int SuperWord::get_iv_adjustment(MemNode* mem_ref) {
 517   SWPointer align_to_ref_p(mem_ref, this);
 518   int offset = align_to_ref_p.offset_in_bytes();
 519   int scale  = align_to_ref_p.scale_in_bytes();
 520   int elt_size = align_to_ref_p.memory_size();
 521   int vw       = vector_width_in_bytes(mem_ref);
 522   assert(vw > 1, "sanity");
 523   int iv_adjustment;
 524   if (scale != 0) {
 525     int stride_sign = (scale * iv_stride()) > 0 ? 1 : -1;
 526     // At least one iteration is executed in pre-loop by default. As result
 527     // several iterations are needed to align memory operations in main-loop even
 528     // if offset is 0.
 529     int iv_adjustment_in_bytes = (stride_sign * vw - (offset % vw));
 530     assert(((ABS(iv_adjustment_in_bytes) % elt_size) == 0),
 531            err_msg_res("(%d) should be divisible by (%d)", iv_adjustment_in_bytes, elt_size));
 532     iv_adjustment = iv_adjustment_in_bytes/elt_size;
 533   } else {
 534     // This memory op is not dependent on iv (scale == 0)
 535     iv_adjustment = 0;
 536   }
 537 
 538 #ifndef PRODUCT
 539   if (TraceSuperWord)
 540     tty->print_cr("\noffset = %d iv_adjust = %d elt_size = %d scale = %d iv_stride = %d vect_size %d",
 541                   offset, iv_adjustment, elt_size, scale, iv_stride(), vw);
 542 #endif
 543   return iv_adjustment;
 544 }
 545 
 546 //---------------------------dependence_graph---------------------------
 547 // Construct dependency graph.
 548 // Add dependence edges to load/store nodes for memory dependence
 549 //    A.out()->DependNode.in(1) and DependNode.out()->B.prec(x)
 550 void SuperWord::dependence_graph() {
 551   // First, assign a dependence node to each memory node
 552   for (int i = 0; i < _block.length(); i++ ) {
 553     Node *n = _block.at(i);
 554     if (n->is_Mem() || n->is_Phi() && n->bottom_type() == Type::MEMORY) {
 555       _dg.make_node(n);
 556     }
 557   }
 558 
 559   // For each memory slice, create the dependences
 560   for (int i = 0; i < _mem_slice_head.length(); i++) {
 561     Node* n      = _mem_slice_head.at(i);
 562     Node* n_tail = _mem_slice_tail.at(i);
 563 
 564     // Get slice in predecessor order (last is first)
 565     mem_slice_preds(n_tail, n, _nlist);
 566 
 567     // Make the slice dependent on the root
 568     DepMem* slice = _dg.dep(n);
 569     _dg.make_edge(_dg.root(), slice);
 570 
 571     // Create a sink for the slice
 572     DepMem* slice_sink = _dg.make_node(NULL);
 573     _dg.make_edge(slice_sink, _dg.tail());
 574 
 575     // Now visit each pair of memory ops, creating the edges
 576     for (int j = _nlist.length() - 1; j >= 0 ; j--) {
 577       Node* s1 = _nlist.at(j);
 578 
 579       // If no dependency yet, use slice
 580       if (_dg.dep(s1)->in_cnt() == 0) {
 581         _dg.make_edge(slice, s1);
 582       }
 583       SWPointer p1(s1->as_Mem(), this);
 584       bool sink_dependent = true;
 585       for (int k = j - 1; k >= 0; k--) {
 586         Node* s2 = _nlist.at(k);
 587         if (s1->is_Load() && s2->is_Load())
 588           continue;
 589         SWPointer p2(s2->as_Mem(), this);
 590 
 591         int cmp = p1.cmp(p2);
 592         if (SuperWordRTDepCheck &&
 593             p1.base() != p2.base() && p1.valid() && p2.valid()) {
 594           // Create a runtime check to disambiguate
 595           OrderedPair pp(p1.base(), p2.base());
 596           _disjoint_ptrs.append_if_missing(pp);
 597         } else if (!SWPointer::not_equal(cmp)) {
 598           // Possibly same address
 599           _dg.make_edge(s1, s2);
 600           sink_dependent = false;
 601         }
 602       }
 603       if (sink_dependent) {
 604         _dg.make_edge(s1, slice_sink);
 605       }
 606     }
 607 #ifndef PRODUCT
 608     if (TraceSuperWord) {
 609       tty->print_cr("\nDependence graph for slice: %d", n->_idx);
 610       for (int q = 0; q < _nlist.length(); q++) {
 611         _dg.print(_nlist.at(q));
 612       }
 613       tty->cr();
 614     }
 615 #endif
 616     _nlist.clear();
 617   }
 618 
 619 #ifndef PRODUCT
 620   if (TraceSuperWord) {
 621     tty->print_cr("\ndisjoint_ptrs: %s", _disjoint_ptrs.length() > 0 ? "" : "NONE");
 622     for (int r = 0; r < _disjoint_ptrs.length(); r++) {
 623       _disjoint_ptrs.at(r).print();
 624       tty->cr();
 625     }
 626     tty->cr();
 627   }
 628 #endif
 629 }
 630 
 631 //---------------------------mem_slice_preds---------------------------
 632 // Return a memory slice (node list) in predecessor order starting at "start"
 633 void SuperWord::mem_slice_preds(Node* start, Node* stop, GrowableArray<Node*> &preds) {
 634   assert(preds.length() == 0, "start empty");
 635   Node* n = start;
 636   Node* prev = NULL;
 637   while (true) {
 638     assert(in_bb(n), "must be in block");
 639     for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
 640       Node* out = n->fast_out(i);
 641       if (out->is_Load()) {
 642         if (in_bb(out)) {
 643           preds.push(out);
 644         }
 645       } else {
 646         // FIXME
 647         if (out->is_MergeMem() && !in_bb(out)) {
 648           // Either unrolling is causing a memory edge not to disappear,
 649           // or need to run igvn.optimize() again before SLP
 650         } else if (out->is_Phi() && out->bottom_type() == Type::MEMORY && !in_bb(out)) {
 651           // Ditto.  Not sure what else to check further.
 652         } else if (out->Opcode() == Op_StoreCM && out->in(MemNode::OopStore) == n) {
 653           // StoreCM has an input edge used as a precedence edge.
 654           // Maybe an issue when oop stores are vectorized.
 655         } else {
 656           assert(out == prev || prev == NULL, "no branches off of store slice");
 657         }
 658       }
 659     }
 660     if (n == stop) break;
 661     preds.push(n);
 662     prev = n;
 663     assert(n->is_Mem(), err_msg_res("unexpected node %s", n->Name()));
 664     n = n->in(MemNode::Memory);
 665   }
 666 }
 667 
 668 //------------------------------stmts_can_pack---------------------------
 669 // Can s1 and s2 be in a pack with s1 immediately preceding s2 and
 670 // s1 aligned at "align"
 671 bool SuperWord::stmts_can_pack(Node* s1, Node* s2, int align) {
 672 
 673   // Do not use superword for non-primitives
 674   BasicType bt1 = velt_basic_type(s1);
 675   BasicType bt2 = velt_basic_type(s2);
 676   if(!is_java_primitive(bt1) || !is_java_primitive(bt2))
 677     return false;
 678   if (Matcher::max_vector_size(bt1) < 2) {
 679     return false; // No vectors for this type
 680   }
 681 
 682   if (isomorphic(s1, s2)) {
 683     if (independent(s1, s2)) {
 684       if (!exists_at(s1, 0) && !exists_at(s2, 1)) {
 685         if (!s1->is_Mem() || are_adjacent_refs(s1, s2)) {
 686           int s1_align = alignment(s1);
 687           int s2_align = alignment(s2);
 688           if (s1_align == top_align || s1_align == align) {
 689             if (s2_align == top_align || s2_align == align + data_size(s1)) {
 690               return true;
 691             }
 692           }
 693         }
 694       }
 695     }
 696   }
 697   return false;
 698 }
 699 
 700 //------------------------------exists_at---------------------------
 701 // Does s exist in a pack at position pos?
 702 bool SuperWord::exists_at(Node* s, uint pos) {
 703   for (int i = 0; i < _packset.length(); i++) {
 704     Node_List* p = _packset.at(i);
 705     if (p->at(pos) == s) {
 706       return true;
 707     }
 708   }
 709   return false;
 710 }
 711 
 712 //------------------------------are_adjacent_refs---------------------------
 713 // Is s1 immediately before s2 in memory?
 714 bool SuperWord::are_adjacent_refs(Node* s1, Node* s2) {
 715   if (!s1->is_Mem() || !s2->is_Mem()) return false;
 716   if (!in_bb(s1)    || !in_bb(s2))    return false;
 717 
 718   // Do not use superword for non-primitives
 719   if (!is_java_primitive(s1->as_Mem()->memory_type()) ||
 720       !is_java_primitive(s2->as_Mem()->memory_type())) {
 721     return false;
 722   }
 723 
 724   // FIXME - co_locate_pack fails on Stores in different mem-slices, so
 725   // only pack memops that are in the same alias set until that's fixed.
 726   if (_phase->C->get_alias_index(s1->as_Mem()->adr_type()) !=
 727       _phase->C->get_alias_index(s2->as_Mem()->adr_type()))
 728     return false;
 729   SWPointer p1(s1->as_Mem(), this);
 730   SWPointer p2(s2->as_Mem(), this);
 731   if (p1.base() != p2.base() || !p1.comparable(p2)) return false;
 732   int diff = p2.offset_in_bytes() - p1.offset_in_bytes();
 733   return diff == data_size(s1);
 734 }
 735 
 736 //------------------------------isomorphic---------------------------
 737 // Are s1 and s2 similar?
 738 bool SuperWord::isomorphic(Node* s1, Node* s2) {
 739   if (s1->Opcode() != s2->Opcode()) return false;
 740   if (s1->req() != s2->req()) return false;
 741   if (s1->in(0) != s2->in(0)) return false;
 742   if (!same_velt_type(s1, s2)) return false;
 743   return true;
 744 }
 745 
 746 //------------------------------independent---------------------------
 747 // Is there no data path from s1 to s2 or s2 to s1?
 748 bool SuperWord::independent(Node* s1, Node* s2) {
 749   //  assert(s1->Opcode() == s2->Opcode(), "check isomorphic first");
 750   int d1 = depth(s1);
 751   int d2 = depth(s2);
 752   if (d1 == d2) return s1 != s2;
 753   Node* deep    = d1 > d2 ? s1 : s2;
 754   Node* shallow = d1 > d2 ? s2 : s1;
 755 
 756   visited_clear();
 757 
 758   return independent_path(shallow, deep);
 759 }
 760 
 761 //------------------------------independent_path------------------------------
 762 // Helper for independent
 763 bool SuperWord::independent_path(Node* shallow, Node* deep, uint dp) {
 764   if (dp >= 1000) return false; // stop deep recursion
 765   visited_set(deep);
 766   int shal_depth = depth(shallow);
 767   assert(shal_depth <= depth(deep), "must be");
 768   for (DepPreds preds(deep, _dg); !preds.done(); preds.next()) {
 769     Node* pred = preds.current();
 770     if (in_bb(pred) && !visited_test(pred)) {
 771       if (shallow == pred) {
 772         return false;
 773       }
 774       if (shal_depth < depth(pred) && !independent_path(shallow, pred, dp+1)) {
 775         return false;
 776       }
 777     }
 778   }
 779   return true;
 780 }
 781 
 782 //------------------------------set_alignment---------------------------
 783 void SuperWord::set_alignment(Node* s1, Node* s2, int align) {
 784   set_alignment(s1, align);
 785   if (align == top_align || align == bottom_align) {
 786     set_alignment(s2, align);
 787   } else {
 788     set_alignment(s2, align + data_size(s1));
 789   }
 790 }
 791 
 792 //------------------------------data_size---------------------------
 793 int SuperWord::data_size(Node* s) {
 794   int bsize = type2aelembytes(velt_basic_type(s));
 795   assert(bsize != 0, "valid size");
 796   return bsize;
 797 }
 798 
 799 //------------------------------extend_packlist---------------------------
 800 // Extend packset by following use->def and def->use links from pack members.
 801 void SuperWord::extend_packlist() {
 802   bool changed;
 803   do {
 804     changed = false;
 805     for (int i = 0; i < _packset.length(); i++) {
 806       Node_List* p = _packset.at(i);
 807       changed |= follow_use_defs(p);
 808       changed |= follow_def_uses(p);
 809     }
 810   } while (changed);
 811 
 812 #ifndef PRODUCT
 813   if (TraceSuperWord) {
 814     tty->print_cr("\nAfter extend_packlist");
 815     print_packset();
 816   }
 817 #endif
 818 }
 819 
 820 //------------------------------follow_use_defs---------------------------
 821 // Extend the packset by visiting operand definitions of nodes in pack p
 822 bool SuperWord::follow_use_defs(Node_List* p) {
 823   assert(p->size() == 2, "just checking");
 824   Node* s1 = p->at(0);
 825   Node* s2 = p->at(1);
 826   assert(s1->req() == s2->req(), "just checking");
 827   assert(alignment(s1) + data_size(s1) == alignment(s2), "just checking");
 828 
 829   if (s1->is_Load()) return false;
 830 
 831   int align = alignment(s1);
 832   bool changed = false;
 833   int start = s1->is_Store() ? MemNode::ValueIn   : 1;
 834   int end   = s1->is_Store() ? MemNode::ValueIn+1 : s1->req();
 835   for (int j = start; j < end; j++) {
 836     Node* t1 = s1->in(j);
 837     Node* t2 = s2->in(j);
 838     if (!in_bb(t1) || !in_bb(t2))
 839       continue;
 840     if (stmts_can_pack(t1, t2, align)) {
 841       if (est_savings(t1, t2) >= 0) {
 842         Node_List* pair = new Node_List();
 843         pair->push(t1);
 844         pair->push(t2);
 845         _packset.append(pair);
 846         set_alignment(t1, t2, align);
 847         changed = true;
 848       }
 849     }
 850   }
 851   return changed;
 852 }
 853 
 854 //------------------------------follow_def_uses---------------------------
 855 // Extend the packset by visiting uses of nodes in pack p
 856 bool SuperWord::follow_def_uses(Node_List* p) {
 857   bool changed = false;
 858   Node* s1 = p->at(0);
 859   Node* s2 = p->at(1);
 860   assert(p->size() == 2, "just checking");
 861   assert(s1->req() == s2->req(), "just checking");
 862   assert(alignment(s1) + data_size(s1) == alignment(s2), "just checking");
 863 
 864   if (s1->is_Store()) return false;
 865 
 866   int align = alignment(s1);
 867   int savings = -1;
 868   Node* u1 = NULL;
 869   Node* u2 = NULL;
 870   for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
 871     Node* t1 = s1->fast_out(i);
 872     if (!in_bb(t1)) continue;
 873     for (DUIterator_Fast jmax, j = s2->fast_outs(jmax); j < jmax; j++) {
 874       Node* t2 = s2->fast_out(j);
 875       if (!in_bb(t2)) continue;
 876       if (!opnd_positions_match(s1, t1, s2, t2))
 877         continue;
 878       if (stmts_can_pack(t1, t2, align)) {
 879         int my_savings = est_savings(t1, t2);
 880         if (my_savings > savings) {
 881           savings = my_savings;
 882           u1 = t1;
 883           u2 = t2;
 884         }
 885       }
 886     }
 887   }
 888   if (savings >= 0) {
 889     Node_List* pair = new Node_List();
 890     pair->push(u1);
 891     pair->push(u2);
 892     _packset.append(pair);
 893     set_alignment(u1, u2, align);
 894     changed = true;
 895   }
 896   return changed;
 897 }
 898 
 899 //---------------------------opnd_positions_match-------------------------
 900 // Is the use of d1 in u1 at the same operand position as d2 in u2?
 901 bool SuperWord::opnd_positions_match(Node* d1, Node* u1, Node* d2, Node* u2) {
 902   uint ct = u1->req();
 903   if (ct != u2->req()) return false;
 904   uint i1 = 0;
 905   uint i2 = 0;
 906   do {
 907     for (i1++; i1 < ct; i1++) if (u1->in(i1) == d1) break;
 908     for (i2++; i2 < ct; i2++) if (u2->in(i2) == d2) break;
 909     if (i1 != i2) {
 910       if ((i1 == (3-i2)) && (u2->is_Add() || u2->is_Mul())) {
 911         // Further analysis relies on operands position matching.
 912         u2->swap_edges(i1, i2);
 913       } else {
 914         return false;
 915       }
 916     }
 917   } while (i1 < ct);
 918   return true;
 919 }
 920 
 921 //------------------------------est_savings---------------------------
 922 // Estimate the savings from executing s1 and s2 as a pack
 923 int SuperWord::est_savings(Node* s1, Node* s2) {
 924   int save_in = 2 - 1; // 2 operations per instruction in packed form
 925 
 926   // inputs
 927   for (uint i = 1; i < s1->req(); i++) {
 928     Node* x1 = s1->in(i);
 929     Node* x2 = s2->in(i);
 930     if (x1 != x2) {
 931       if (are_adjacent_refs(x1, x2)) {
 932         save_in += adjacent_profit(x1, x2);
 933       } else if (!in_packset(x1, x2)) {
 934         save_in -= pack_cost(2);
 935       } else {
 936         save_in += unpack_cost(2);
 937       }
 938     }
 939   }
 940 
 941   // uses of result
 942   uint ct = 0;
 943   int save_use = 0;
 944   for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
 945     Node* s1_use = s1->fast_out(i);
 946     for (int j = 0; j < _packset.length(); j++) {
 947       Node_List* p = _packset.at(j);
 948       if (p->at(0) == s1_use) {
 949         for (DUIterator_Fast kmax, k = s2->fast_outs(kmax); k < kmax; k++) {
 950           Node* s2_use = s2->fast_out(k);
 951           if (p->at(p->size()-1) == s2_use) {
 952             ct++;
 953             if (are_adjacent_refs(s1_use, s2_use)) {
 954               save_use += adjacent_profit(s1_use, s2_use);
 955             }
 956           }
 957         }
 958       }
 959     }
 960   }
 961 
 962   if (ct < s1->outcnt()) save_use += unpack_cost(1);
 963   if (ct < s2->outcnt()) save_use += unpack_cost(1);
 964 
 965   return MAX2(save_in, save_use);
 966 }
 967 
 968 //------------------------------costs---------------------------
 969 int SuperWord::adjacent_profit(Node* s1, Node* s2) { return 2; }
 970 int SuperWord::pack_cost(int ct)   { return ct; }
 971 int SuperWord::unpack_cost(int ct) { return ct; }
 972 
 973 //------------------------------combine_packs---------------------------
 974 // Combine packs A and B with A.last == B.first into A.first..,A.last,B.second,..B.last
 975 void SuperWord::combine_packs() {
 976   bool changed = true;
 977   // Combine packs regardless max vector size.
 978   while (changed) {
 979     changed = false;
 980     for (int i = 0; i < _packset.length(); i++) {
 981       Node_List* p1 = _packset.at(i);
 982       if (p1 == NULL) continue;
 983       for (int j = 0; j < _packset.length(); j++) {
 984         Node_List* p2 = _packset.at(j);
 985         if (p2 == NULL) continue;
 986         if (i == j) continue;
 987         if (p1->at(p1->size()-1) == p2->at(0)) {
 988           for (uint k = 1; k < p2->size(); k++) {
 989             p1->push(p2->at(k));
 990           }
 991           _packset.at_put(j, NULL);
 992           changed = true;
 993         }
 994       }
 995     }
 996   }
 997 
 998   // Split packs which have size greater then max vector size.
 999   for (int i = 0; i < _packset.length(); i++) {
1000     Node_List* p1 = _packset.at(i);
1001     if (p1 != NULL) {
1002       BasicType bt = velt_basic_type(p1->at(0));
1003       uint max_vlen = Matcher::max_vector_size(bt); // Max elements in vector
1004       assert(is_power_of_2(max_vlen), "sanity");
1005       uint psize = p1->size();
1006       if (!is_power_of_2(psize)) {
1007         // Skip pack which can't be vector.
1008         // case1: for(...) { a[i] = i; }    elements values are different (i+x)
1009         // case2: for(...) { a[i] = b[i+1]; }  can't align both, load and store
1010         _packset.at_put(i, NULL);
1011         continue;
1012       }
1013       if (psize > max_vlen) {
1014         Node_List* pack = new Node_List();
1015         for (uint j = 0; j < psize; j++) {
1016           pack->push(p1->at(j));
1017           if (pack->size() >= max_vlen) {
1018             assert(is_power_of_2(pack->size()), "sanity");
1019             _packset.append(pack);
1020             pack = new Node_List();
1021           }
1022         }
1023         _packset.at_put(i, NULL);
1024       }
1025     }
1026   }
1027 
1028   // Compress list.
1029   for (int i = _packset.length() - 1; i >= 0; i--) {
1030     Node_List* p1 = _packset.at(i);
1031     if (p1 == NULL) {
1032       _packset.remove_at(i);
1033     }
1034   }
1035 
1036 #ifndef PRODUCT
1037   if (TraceSuperWord) {
1038     tty->print_cr("\nAfter combine_packs");
1039     print_packset();
1040   }
1041 #endif
1042 }
1043 
1044 //-----------------------------construct_my_pack_map--------------------------
1045 // Construct the map from nodes to packs.  Only valid after the
1046 // point where a node is only in one pack (after combine_packs).
1047 void SuperWord::construct_my_pack_map() {
1048   Node_List* rslt = NULL;
1049   for (int i = 0; i < _packset.length(); i++) {
1050     Node_List* p = _packset.at(i);
1051     for (uint j = 0; j < p->size(); j++) {
1052       Node* s = p->at(j);
1053       assert(my_pack(s) == NULL, "only in one pack");
1054       set_my_pack(s, p);
1055     }
1056   }
1057 }
1058 
1059 //------------------------------filter_packs---------------------------
1060 // Remove packs that are not implemented or not profitable.
1061 void SuperWord::filter_packs() {
1062 
1063   // Remove packs that are not implemented
1064   for (int i = _packset.length() - 1; i >= 0; i--) {
1065     Node_List* pk = _packset.at(i);
1066     bool impl = implemented(pk);
1067     if (!impl) {
1068 #ifndef PRODUCT
1069       if (TraceSuperWord && Verbose) {
1070         tty->print_cr("Unimplemented");
1071         pk->at(0)->dump();
1072       }
1073 #endif
1074       remove_pack_at(i);
1075     }
1076   }
1077 
1078   // Remove packs that are not profitable
1079   bool changed;
1080   do {
1081     changed = false;
1082     for (int i = _packset.length() - 1; i >= 0; i--) {
1083       Node_List* pk = _packset.at(i);
1084       bool prof = profitable(pk);
1085       if (!prof) {
1086 #ifndef PRODUCT
1087         if (TraceSuperWord && Verbose) {
1088           tty->print_cr("Unprofitable");
1089           pk->at(0)->dump();
1090         }
1091 #endif
1092         remove_pack_at(i);
1093         changed = true;
1094       }
1095     }
1096   } while (changed);
1097 
1098 #ifndef PRODUCT
1099   if (TraceSuperWord) {
1100     tty->print_cr("\nAfter filter_packs");
1101     print_packset();
1102     tty->cr();
1103   }
1104 #endif
1105 }
1106 
1107 //------------------------------implemented---------------------------
1108 // Can code be generated for pack p?
1109 bool SuperWord::implemented(Node_List* p) {
1110   Node* p0 = p->at(0);
1111   return VectorNode::implemented(p0->Opcode(), p->size(), velt_basic_type(p0));
1112 }
1113 
1114 //------------------------------same_inputs--------------------------
1115 // For pack p, are all idx operands the same?
1116 static bool same_inputs(Node_List* p, int idx) {
1117   Node* p0 = p->at(0);
1118   uint vlen = p->size();
1119   Node* p0_def = p0->in(idx);
1120   for (uint i = 1; i < vlen; i++) {
1121     Node* pi = p->at(i);
1122     Node* pi_def = pi->in(idx);
1123     if (p0_def != pi_def)
1124       return false;
1125   }
1126   return true;
1127 }
1128 
1129 //------------------------------profitable---------------------------
1130 // For pack p, are all operands and all uses (with in the block) vector?
1131 bool SuperWord::profitable(Node_List* p) {
1132   Node* p0 = p->at(0);
1133   uint start, end;
1134   VectorNode::vector_operands(p0, &start, &end);
1135 
1136   // Return false if some inputs are not vectors or vectors with different
1137   // size or alignment.
1138   // Also, for now, return false if not scalar promotion case when inputs are
1139   // the same. Later, implement PackNode and allow differing, non-vector inputs
1140   // (maybe just the ones from outside the block.)
1141   for (uint i = start; i < end; i++) {
1142     if (!is_vector_use(p0, i))
1143       return false;
1144   }
1145   if (VectorNode::is_shift(p0)) {
1146     // For now, return false if shift count is vector or not scalar promotion
1147     // case (different shift counts) because it is not supported yet.
1148     Node* cnt = p0->in(2);
1149     Node_List* cnt_pk = my_pack(cnt);
1150     if (cnt_pk != NULL)
1151       return false;
1152     if (!same_inputs(p, 2))
1153       return false;
1154   }
1155   if (!p0->is_Store()) {
1156     // For now, return false if not all uses are vector.
1157     // Later, implement ExtractNode and allow non-vector uses (maybe
1158     // just the ones outside the block.)
1159     for (uint i = 0; i < p->size(); i++) {
1160       Node* def = p->at(i);
1161       for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
1162         Node* use = def->fast_out(j);
1163         for (uint k = 0; k < use->req(); k++) {
1164           Node* n = use->in(k);
1165           if (def == n) {
1166             if (!is_vector_use(use, k)) {
1167               return false;
1168             }
1169           }
1170         }
1171       }
1172     }
1173   }
1174   return true;
1175 }
1176 
1177 //------------------------------schedule---------------------------
1178 // Adjust the memory graph for the packed operations
1179 void SuperWord::schedule() {
1180 
1181   // Co-locate in the memory graph the members of each memory pack
1182   for (int i = 0; i < _packset.length(); i++) {
1183     co_locate_pack(_packset.at(i));
1184   }
1185 }
1186 
1187 //-------------------------------remove_and_insert-------------------
1188 // Remove "current" from its current position in the memory graph and insert
1189 // it after the appropriate insertion point (lip or uip).
1190 void SuperWord::remove_and_insert(MemNode *current, MemNode *prev, MemNode *lip,
1191                                   Node *uip, Unique_Node_List &sched_before) {
1192   Node* my_mem = current->in(MemNode::Memory);
1193   bool sched_up = sched_before.member(current);
1194 
1195   // remove current_store from its current position in the memmory graph
1196   for (DUIterator i = current->outs(); current->has_out(i); i++) {
1197     Node* use = current->out(i);
1198     if (use->is_Mem()) {
1199       assert(use->in(MemNode::Memory) == current, "must be");
1200       if (use == prev) { // connect prev to my_mem
1201           _igvn.replace_input_of(use, MemNode::Memory, my_mem);
1202           --i; //deleted this edge; rescan position
1203       } else if (sched_before.member(use)) {
1204         if (!sched_up) { // Will be moved together with current
1205           _igvn.replace_input_of(use, MemNode::Memory, uip);
1206           --i; //deleted this edge; rescan position
1207         }
1208       } else {
1209         if (sched_up) { // Will be moved together with current
1210           _igvn.replace_input_of(use, MemNode::Memory, lip);
1211           --i; //deleted this edge; rescan position
1212         }
1213       }
1214     }
1215   }
1216 
1217   Node *insert_pt =  sched_up ?  uip : lip;
1218 
1219   // all uses of insert_pt's memory state should use current's instead
1220   for (DUIterator i = insert_pt->outs(); insert_pt->has_out(i); i++) {
1221     Node* use = insert_pt->out(i);
1222     if (use->is_Mem()) {
1223       assert(use->in(MemNode::Memory) == insert_pt, "must be");
1224       _igvn.replace_input_of(use, MemNode::Memory, current);
1225       --i; //deleted this edge; rescan position
1226     } else if (!sched_up && use->is_Phi() && use->bottom_type() == Type::MEMORY) {
1227       uint pos; //lip (lower insert point) must be the last one in the memory slice
1228       for (pos=1; pos < use->req(); pos++) {
1229         if (use->in(pos) == insert_pt) break;
1230       }
1231       _igvn.replace_input_of(use, pos, current);
1232       --i;
1233     }
1234   }
1235 
1236   //connect current to insert_pt
1237   _igvn.replace_input_of(current, MemNode::Memory, insert_pt);
1238 }
1239 
1240 //------------------------------co_locate_pack----------------------------------
1241 // To schedule a store pack, we need to move any sandwiched memory ops either before
1242 // or after the pack, based upon dependence information:
1243 // (1) If any store in the pack depends on the sandwiched memory op, the
1244 //     sandwiched memory op must be scheduled BEFORE the pack;
1245 // (2) If a sandwiched memory op depends on any store in the pack, the
1246 //     sandwiched memory op must be scheduled AFTER the pack;
1247 // (3) If a sandwiched memory op (say, memA) depends on another sandwiched
1248 //     memory op (say memB), memB must be scheduled before memA. So, if memA is
1249 //     scheduled before the pack, memB must also be scheduled before the pack;
1250 // (4) If there is no dependence restriction for a sandwiched memory op, we simply
1251 //     schedule this store AFTER the pack
1252 // (5) We know there is no dependence cycle, so there in no other case;
1253 // (6) Finally, all memory ops in another single pack should be moved in the same direction.
1254 //
1255 // To schedule a load pack, we use the memory state of either the first or the last load in
1256 // the pack, based on the dependence constraint.
1257 void SuperWord::co_locate_pack(Node_List* pk) {
1258   if (pk->at(0)->is_Store()) {
1259     MemNode* first     = executed_first(pk)->as_Mem();
1260     MemNode* last      = executed_last(pk)->as_Mem();
1261     Unique_Node_List schedule_before_pack;
1262     Unique_Node_List memops;
1263 
1264     MemNode* current   = last->in(MemNode::Memory)->as_Mem();
1265     MemNode* previous  = last;
1266     while (true) {
1267       assert(in_bb(current), "stay in block");
1268       memops.push(previous);
1269       for (DUIterator i = current->outs(); current->has_out(i); i++) {
1270         Node* use = current->out(i);
1271         if (use->is_Mem() && use != previous)
1272           memops.push(use);
1273       }
1274       if (current == first) break;
1275       previous = current;
1276       current  = current->in(MemNode::Memory)->as_Mem();
1277     }
1278 
1279     // determine which memory operations should be scheduled before the pack
1280     for (uint i = 1; i < memops.size(); i++) {
1281       Node *s1 = memops.at(i);
1282       if (!in_pack(s1, pk) && !schedule_before_pack.member(s1)) {
1283         for (uint j = 0; j< i; j++) {
1284           Node *s2 = memops.at(j);
1285           if (!independent(s1, s2)) {
1286             if (in_pack(s2, pk) || schedule_before_pack.member(s2)) {
1287               schedule_before_pack.push(s1); // s1 must be scheduled before
1288               Node_List* mem_pk = my_pack(s1);
1289               if (mem_pk != NULL) {
1290                 for (uint ii = 0; ii < mem_pk->size(); ii++) {
1291                   Node* s = mem_pk->at(ii);  // follow partner
1292                   if (memops.member(s) && !schedule_before_pack.member(s))
1293                     schedule_before_pack.push(s);
1294                 }
1295               }
1296               break;
1297             }
1298           }
1299         }
1300       }
1301     }
1302 
1303     Node*    upper_insert_pt = first->in(MemNode::Memory);
1304     // Following code moves loads connected to upper_insert_pt below aliased stores.
1305     // Collect such loads here and reconnect them back to upper_insert_pt later.
1306     memops.clear();
1307     for (DUIterator i = upper_insert_pt->outs(); upper_insert_pt->has_out(i); i++) {
1308       Node* use = upper_insert_pt->out(i);
1309       if (use->is_Mem() && !use->is_Store()) {
1310         memops.push(use);
1311       }
1312     }
1313 
1314     MemNode* lower_insert_pt = last;
1315     previous                 = last; //previous store in pk
1316     current                  = last->in(MemNode::Memory)->as_Mem();
1317 
1318     // start scheduling from "last" to "first"
1319     while (true) {
1320       assert(in_bb(current), "stay in block");
1321       assert(in_pack(previous, pk), "previous stays in pack");
1322       Node* my_mem = current->in(MemNode::Memory);
1323 
1324       if (in_pack(current, pk)) {
1325         // Forward users of my memory state (except "previous) to my input memory state
1326         for (DUIterator i = current->outs(); current->has_out(i); i++) {
1327           Node* use = current->out(i);
1328           if (use->is_Mem() && use != previous) {
1329             assert(use->in(MemNode::Memory) == current, "must be");
1330             if (schedule_before_pack.member(use)) {
1331               _igvn.replace_input_of(use, MemNode::Memory, upper_insert_pt);
1332             } else {
1333               _igvn.replace_input_of(use, MemNode::Memory, lower_insert_pt);
1334             }
1335             --i; // deleted this edge; rescan position
1336           }
1337         }
1338         previous = current;
1339       } else { // !in_pack(current, pk) ==> a sandwiched store
1340         remove_and_insert(current, previous, lower_insert_pt, upper_insert_pt, schedule_before_pack);
1341       }
1342 
1343       if (current == first) break;
1344       current = my_mem->as_Mem();
1345     } // end while
1346 
1347     // Reconnect loads back to upper_insert_pt.
1348     for (uint i = 0; i < memops.size(); i++) {
1349       Node *ld = memops.at(i);
1350       if (ld->in(MemNode::Memory) != upper_insert_pt) {
1351         _igvn.replace_input_of(ld, MemNode::Memory, upper_insert_pt);
1352       }
1353     }
1354   } else if (pk->at(0)->is_Load()) { //load
1355     // all loads in the pack should have the same memory state. By default,
1356     // we use the memory state of the last load. However, if any load could
1357     // not be moved down due to the dependence constraint, we use the memory
1358     // state of the first load.
1359     Node* last_mem  = executed_last(pk)->in(MemNode::Memory);
1360     Node* first_mem = executed_first(pk)->in(MemNode::Memory);
1361     bool schedule_last = true;
1362     for (uint i = 0; i < pk->size(); i++) {
1363       Node* ld = pk->at(i);
1364       for (Node* current = last_mem; current != ld->in(MemNode::Memory);
1365            current=current->in(MemNode::Memory)) {
1366         assert(current != first_mem, "corrupted memory graph");
1367         if(current->is_Mem() && !independent(current, ld)){
1368           schedule_last = false; // a later store depends on this load
1369           break;
1370         }
1371       }
1372     }
1373 
1374     Node* mem_input = schedule_last ? last_mem : first_mem;
1375     _igvn.hash_delete(mem_input);
1376     // Give each load the same memory state
1377     for (uint i = 0; i < pk->size(); i++) {
1378       LoadNode* ld = pk->at(i)->as_Load();
1379       _igvn.replace_input_of(ld, MemNode::Memory, mem_input);
1380     }
1381   }
1382 }
1383 
1384 //------------------------------output---------------------------
1385 // Convert packs into vector node operations
1386 void SuperWord::output() {
1387   if (_packset.length() == 0) return;
1388 
1389 #ifndef PRODUCT
1390   if (TraceLoopOpts) {
1391     tty->print("SuperWord    ");
1392     lpt()->dump_head();
1393   }
1394 #endif
1395 
1396   // MUST ENSURE main loop's initial value is properly aligned:
1397   //  (iv_initial_value + min_iv_offset) % vector_width_in_bytes() == 0
1398 
1399   align_initial_loop_index(align_to_ref());
1400 
1401   // Insert extract (unpack) operations for scalar uses
1402   for (int i = 0; i < _packset.length(); i++) {
1403     insert_extracts(_packset.at(i));
1404   }
1405 
1406   Compile* C = _phase->C;
1407   uint max_vlen_in_bytes = 0;
1408   for (int i = 0; i < _block.length(); i++) {
1409     Node* n = _block.at(i);
1410     Node_List* p = my_pack(n);
1411     if (p && n == executed_last(p)) {
1412       uint vlen = p->size();
1413       uint vlen_in_bytes = 0;
1414       Node* vn = NULL;
1415       Node* low_adr = p->at(0);
1416       Node* first   = executed_first(p);
1417       int   opc = n->Opcode();
1418       if (n->is_Load()) {
1419         Node* ctl = n->in(MemNode::Control);
1420         Node* mem = first->in(MemNode::Memory);
1421         SWPointer p1(n->as_Mem(), this);
1422         // Identify the memory dependency for the new loadVector node by
1423         // walking up through memory chain.
1424         // This is done to give flexibility to the new loadVector node so that
1425         // it can move above independent storeVector nodes.
1426         while (mem->is_StoreVector()) {
1427           SWPointer p2(mem->as_Mem(), this);
1428           int cmp = p1.cmp(p2);
1429           if (SWPointer::not_equal(cmp) || !SWPointer::comparable(cmp)) {
1430             mem = mem->in(MemNode::Memory);
1431           } else {
1432             break; // dependent memory
1433           }
1434         }
1435         Node* adr = low_adr->in(MemNode::Address);
1436         const TypePtr* atyp = n->adr_type();
1437         vn = LoadVectorNode::make(C, opc, ctl, mem, adr, atyp, vlen, velt_basic_type(n), control_dependency(p));
1438         vlen_in_bytes = vn->as_LoadVector()->memory_size();
1439       } else if (n->is_Store()) {
1440         // Promote value to be stored to vector
1441         Node* val = vector_opd(p, MemNode::ValueIn);
1442         Node* ctl = n->in(MemNode::Control);
1443         Node* mem = first->in(MemNode::Memory);
1444         Node* adr = low_adr->in(MemNode::Address);
1445         const TypePtr* atyp = n->adr_type();
1446         vn = StoreVectorNode::make(C, opc, ctl, mem, adr, atyp, val, vlen);
1447         vlen_in_bytes = vn->as_StoreVector()->memory_size();
1448       } else if (n->req() == 3) {
1449         // Promote operands to vector
1450         Node* in1 = vector_opd(p, 1);
1451         Node* in2 = vector_opd(p, 2);
1452         if (VectorNode::is_invariant_vector(in1) && (n->is_Add() || n->is_Mul())) {
1453           // Move invariant vector input into second position to avoid register spilling.
1454           Node* tmp = in1;
1455           in1 = in2;
1456           in2 = tmp;
1457         }
1458         vn = VectorNode::make(C, opc, in1, in2, vlen, velt_basic_type(n));
1459         vlen_in_bytes = vn->as_Vector()->length_in_bytes();
1460       } else {
1461         ShouldNotReachHere();
1462       }
1463       assert(vn != NULL, "sanity");
1464       _igvn.register_new_node_with_optimizer(vn);
1465       _phase->set_ctrl(vn, _phase->get_ctrl(p->at(0)));
1466       for (uint j = 0; j < p->size(); j++) {
1467         Node* pm = p->at(j);
1468         _igvn.replace_node(pm, vn);
1469       }
1470       _igvn._worklist.push(vn);
1471 
1472       if (vlen_in_bytes > max_vlen_in_bytes) {
1473         max_vlen_in_bytes = vlen_in_bytes;
1474       }
1475 #ifdef ASSERT
1476       if (TraceNewVectors) {
1477         tty->print("new Vector node: ");
1478         vn->dump();
1479       }
1480 #endif
1481     }
1482   }
1483   C->set_max_vector_size(max_vlen_in_bytes);
1484 }
1485 
1486 //------------------------------vector_opd---------------------------
1487 // Create a vector operand for the nodes in pack p for operand: in(opd_idx)
1488 Node* SuperWord::vector_opd(Node_List* p, int opd_idx) {
1489   Node* p0 = p->at(0);
1490   uint vlen = p->size();
1491   Node* opd = p0->in(opd_idx);
1492 
1493   if (same_inputs(p, opd_idx)) {
1494     if (opd->is_Vector() || opd->is_LoadVector()) {
1495       assert(((opd_idx != 2) || !VectorNode::is_shift(p0)), "shift's count can't be vector");
1496       return opd; // input is matching vector
1497     }
1498     if ((opd_idx == 2) && VectorNode::is_shift(p0)) {
1499       Compile* C = _phase->C;
1500       Node* cnt = opd;
1501       // Vector instructions do not mask shift count, do it here.
1502       juint mask = (p0->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
1503       const TypeInt* t = opd->find_int_type();
1504       if (t != NULL && t->is_con()) {
1505         juint shift = t->get_con();
1506         if (shift > mask) { // Unsigned cmp
1507           cnt = ConNode::make(C, TypeInt::make(shift & mask));
1508         }
1509       } else {
1510         if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
1511           cnt = ConNode::make(C, TypeInt::make(mask));
1512           _igvn.register_new_node_with_optimizer(cnt);
1513           cnt = new (C) AndINode(opd, cnt);
1514           _igvn.register_new_node_with_optimizer(cnt);
1515           _phase->set_ctrl(cnt, _phase->get_ctrl(opd));
1516         }
1517         assert(opd->bottom_type()->isa_int(), "int type only");
1518         // Move non constant shift count into vector register.
1519         cnt = VectorNode::shift_count(C, p0, cnt, vlen, velt_basic_type(p0));
1520       }
1521       if (cnt != opd) {
1522         _igvn.register_new_node_with_optimizer(cnt);
1523         _phase->set_ctrl(cnt, _phase->get_ctrl(opd));
1524       }
1525       return cnt;
1526     }
1527     assert(!opd->is_StoreVector(), "such vector is not expected here");
1528     // Convert scalar input to vector with the same number of elements as
1529     // p0's vector. Use p0's type because size of operand's container in
1530     // vector should match p0's size regardless operand's size.
1531     const Type* p0_t = velt_type(p0);
1532     VectorNode* vn = VectorNode::scalar2vector(_phase->C, opd, vlen, p0_t);
1533 
1534     _igvn.register_new_node_with_optimizer(vn);
1535     _phase->set_ctrl(vn, _phase->get_ctrl(opd));
1536 #ifdef ASSERT
1537     if (TraceNewVectors) {
1538       tty->print("new Vector node: ");
1539       vn->dump();
1540     }
1541 #endif
1542     return vn;
1543   }
1544 
1545   // Insert pack operation
1546   BasicType bt = velt_basic_type(p0);
1547   PackNode* pk = PackNode::make(_phase->C, opd, vlen, bt);
1548   DEBUG_ONLY( const BasicType opd_bt = opd->bottom_type()->basic_type(); )
1549 
1550   for (uint i = 1; i < vlen; i++) {
1551     Node* pi = p->at(i);
1552     Node* in = pi->in(opd_idx);
1553     assert(my_pack(in) == NULL, "Should already have been unpacked");
1554     assert(opd_bt == in->bottom_type()->basic_type(), "all same type");
1555     pk->add_opd(in);
1556   }
1557   _igvn.register_new_node_with_optimizer(pk);
1558   _phase->set_ctrl(pk, _phase->get_ctrl(opd));
1559 #ifdef ASSERT
1560   if (TraceNewVectors) {
1561     tty->print("new Vector node: ");
1562     pk->dump();
1563   }
1564 #endif
1565   return pk;
1566 }
1567 
1568 //------------------------------insert_extracts---------------------------
1569 // If a use of pack p is not a vector use, then replace the
1570 // use with an extract operation.
1571 void SuperWord::insert_extracts(Node_List* p) {
1572   if (p->at(0)->is_Store()) return;
1573   assert(_n_idx_list.is_empty(), "empty (node,index) list");
1574 
1575   // Inspect each use of each pack member.  For each use that is
1576   // not a vector use, replace the use with an extract operation.
1577 
1578   for (uint i = 0; i < p->size(); i++) {
1579     Node* def = p->at(i);
1580     for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
1581       Node* use = def->fast_out(j);
1582       for (uint k = 0; k < use->req(); k++) {
1583         Node* n = use->in(k);
1584         if (def == n) {
1585           if (!is_vector_use(use, k)) {
1586             _n_idx_list.push(use, k);
1587           }
1588         }
1589       }
1590     }
1591   }
1592 
1593   while (_n_idx_list.is_nonempty()) {
1594     Node* use = _n_idx_list.node();
1595     int   idx = _n_idx_list.index();
1596     _n_idx_list.pop();
1597     Node* def = use->in(idx);
1598 
1599     // Insert extract operation
1600     _igvn.hash_delete(def);
1601     int def_pos = alignment(def) / data_size(def);
1602 
1603     Node* ex = ExtractNode::make(_phase->C, def, def_pos, velt_basic_type(def));
1604     _igvn.register_new_node_with_optimizer(ex);
1605     _phase->set_ctrl(ex, _phase->get_ctrl(def));
1606     _igvn.replace_input_of(use, idx, ex);
1607     _igvn._worklist.push(def);
1608 
1609     bb_insert_after(ex, bb_idx(def));
1610     set_velt_type(ex, velt_type(def));
1611   }
1612 }
1613 
1614 //------------------------------is_vector_use---------------------------
1615 // Is use->in(u_idx) a vector use?
1616 bool SuperWord::is_vector_use(Node* use, int u_idx) {
1617   Node_List* u_pk = my_pack(use);
1618   if (u_pk == NULL) return false;
1619   Node* def = use->in(u_idx);
1620   Node_List* d_pk = my_pack(def);
1621   if (d_pk == NULL) {
1622     // check for scalar promotion
1623     Node* n = u_pk->at(0)->in(u_idx);
1624     for (uint i = 1; i < u_pk->size(); i++) {
1625       if (u_pk->at(i)->in(u_idx) != n) return false;
1626     }
1627     return true;
1628   }
1629   if (u_pk->size() != d_pk->size())
1630     return false;
1631   for (uint i = 0; i < u_pk->size(); i++) {
1632     Node* ui = u_pk->at(i);
1633     Node* di = d_pk->at(i);
1634     if (ui->in(u_idx) != di || alignment(ui) != alignment(di))
1635       return false;
1636   }
1637   return true;
1638 }
1639 
1640 //------------------------------construct_bb---------------------------
1641 // Construct reverse postorder list of block members
1642 bool SuperWord::construct_bb() {
1643   Node* entry = bb();
1644 
1645   assert(_stk.length() == 0,            "stk is empty");
1646   assert(_block.length() == 0,          "block is empty");
1647   assert(_data_entry.length() == 0,     "data_entry is empty");
1648   assert(_mem_slice_head.length() == 0, "mem_slice_head is empty");
1649   assert(_mem_slice_tail.length() == 0, "mem_slice_tail is empty");
1650 
1651   // Find non-control nodes with no inputs from within block,
1652   // create a temporary map from node _idx to bb_idx for use
1653   // by the visited and post_visited sets,
1654   // and count number of nodes in block.
1655   int bb_ct = 0;
1656   for (uint i = 0; i < lpt()->_body.size(); i++ ) {
1657     Node *n = lpt()->_body.at(i);
1658     set_bb_idx(n, i); // Create a temporary map
1659     if (in_bb(n)) {
1660       if (n->is_LoadStore() || n->is_MergeMem() ||
1661           (n->is_Proj() && !n->as_Proj()->is_CFG())) {
1662         // Bailout if the loop has LoadStore, MergeMem or data Proj
1663         // nodes. Superword optimization does not work with them.
1664         return false;
1665       }
1666       bb_ct++;
1667       if (!n->is_CFG()) {
1668         bool found = false;
1669         for (uint j = 0; j < n->req(); j++) {
1670           Node* def = n->in(j);
1671           if (def && in_bb(def)) {
1672             found = true;
1673             break;
1674           }
1675         }
1676         if (!found) {
1677           assert(n != entry, "can't be entry");
1678           _data_entry.push(n);
1679         }
1680       }
1681     }
1682   }
1683 
1684   // Find memory slices (head and tail)
1685   for (DUIterator_Fast imax, i = lp()->fast_outs(imax); i < imax; i++) {
1686     Node *n = lp()->fast_out(i);
1687     if (in_bb(n) && (n->is_Phi() && n->bottom_type() == Type::MEMORY)) {
1688       Node* n_tail  = n->in(LoopNode::LoopBackControl);
1689       if (n_tail != n->in(LoopNode::EntryControl)) {
1690         if (!n_tail->is_Mem()) {
1691           assert(n_tail->is_Mem(), err_msg_res("unexpected node for memory slice: %s", n_tail->Name()));
1692           return false; // Bailout
1693         }
1694         _mem_slice_head.push(n);
1695         _mem_slice_tail.push(n_tail);
1696       }
1697     }
1698   }
1699 
1700   // Create an RPO list of nodes in block
1701 
1702   visited_clear();
1703   post_visited_clear();
1704 
1705   // Push all non-control nodes with no inputs from within block, then control entry
1706   for (int j = 0; j < _data_entry.length(); j++) {
1707     Node* n = _data_entry.at(j);
1708     visited_set(n);
1709     _stk.push(n);
1710   }
1711   visited_set(entry);
1712   _stk.push(entry);
1713 
1714   // Do a depth first walk over out edges
1715   int rpo_idx = bb_ct - 1;
1716   int size;
1717   while ((size = _stk.length()) > 0) {
1718     Node* n = _stk.top(); // Leave node on stack
1719     if (!visited_test_set(n)) {
1720       // forward arc in graph
1721     } else if (!post_visited_test(n)) {
1722       // cross or back arc
1723       for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1724         Node *use = n->fast_out(i);
1725         if (in_bb(use) && !visited_test(use) &&
1726             // Don't go around backedge
1727             (!use->is_Phi() || n == entry)) {
1728           _stk.push(use);
1729         }
1730       }
1731       if (_stk.length() == size) {
1732         // There were no additional uses, post visit node now
1733         _stk.pop(); // Remove node from stack
1734         assert(rpo_idx >= 0, "");
1735         _block.at_put_grow(rpo_idx, n);
1736         rpo_idx--;
1737         post_visited_set(n);
1738         assert(rpo_idx >= 0 || _stk.is_empty(), "");
1739       }
1740     } else {
1741       _stk.pop(); // Remove post-visited node from stack
1742     }
1743   }
1744 
1745   // Create real map of block indices for nodes
1746   for (int j = 0; j < _block.length(); j++) {
1747     Node* n = _block.at(j);
1748     set_bb_idx(n, j);
1749   }
1750 
1751   initialize_bb(); // Ensure extra info is allocated.
1752 
1753 #ifndef PRODUCT
1754   if (TraceSuperWord) {
1755     print_bb();
1756     tty->print_cr("\ndata entry nodes: %s", _data_entry.length() > 0 ? "" : "NONE");
1757     for (int m = 0; m < _data_entry.length(); m++) {
1758       tty->print("%3d ", m);
1759       _data_entry.at(m)->dump();
1760     }
1761     tty->print_cr("\nmemory slices: %s", _mem_slice_head.length() > 0 ? "" : "NONE");
1762     for (int m = 0; m < _mem_slice_head.length(); m++) {
1763       tty->print("%3d ", m); _mem_slice_head.at(m)->dump();
1764       tty->print("    ");    _mem_slice_tail.at(m)->dump();
1765     }
1766   }
1767 #endif
1768   assert(rpo_idx == -1 && bb_ct == _block.length(), "all block members found");
1769   return (_mem_slice_head.length() > 0) || (_data_entry.length() > 0);
1770 }
1771 
1772 //------------------------------initialize_bb---------------------------
1773 // Initialize per node info
1774 void SuperWord::initialize_bb() {
1775   Node* last = _block.at(_block.length() - 1);
1776   grow_node_info(bb_idx(last));
1777 }
1778 
1779 //------------------------------bb_insert_after---------------------------
1780 // Insert n into block after pos
1781 void SuperWord::bb_insert_after(Node* n, int pos) {
1782   int n_pos = pos + 1;
1783   // Make room
1784   for (int i = _block.length() - 1; i >= n_pos; i--) {
1785     _block.at_put_grow(i+1, _block.at(i));
1786   }
1787   for (int j = _node_info.length() - 1; j >= n_pos; j--) {
1788     _node_info.at_put_grow(j+1, _node_info.at(j));
1789   }
1790   // Set value
1791   _block.at_put_grow(n_pos, n);
1792   _node_info.at_put_grow(n_pos, SWNodeInfo::initial);
1793   // Adjust map from node->_idx to _block index
1794   for (int i = n_pos; i < _block.length(); i++) {
1795     set_bb_idx(_block.at(i), i);
1796   }
1797 }
1798 
1799 //------------------------------compute_max_depth---------------------------
1800 // Compute max depth for expressions from beginning of block
1801 // Use to prune search paths during test for independence.
1802 void SuperWord::compute_max_depth() {
1803   int ct = 0;
1804   bool again;
1805   do {
1806     again = false;
1807     for (int i = 0; i < _block.length(); i++) {
1808       Node* n = _block.at(i);
1809       if (!n->is_Phi()) {
1810         int d_orig = depth(n);
1811         int d_in   = 0;
1812         for (DepPreds preds(n, _dg); !preds.done(); preds.next()) {
1813           Node* pred = preds.current();
1814           if (in_bb(pred)) {
1815             d_in = MAX2(d_in, depth(pred));
1816           }
1817         }
1818         if (d_in + 1 != d_orig) {
1819           set_depth(n, d_in + 1);
1820           again = true;
1821         }
1822       }
1823     }
1824     ct++;
1825   } while (again);
1826 #ifndef PRODUCT
1827   if (TraceSuperWord && Verbose)
1828     tty->print_cr("compute_max_depth iterated: %d times", ct);
1829 #endif
1830 }
1831 
1832 //-------------------------compute_vector_element_type-----------------------
1833 // Compute necessary vector element type for expressions
1834 // This propagates backwards a narrower integer type when the
1835 // upper bits of the value are not needed.
1836 // Example:  char a,b,c;  a = b + c;
1837 // Normally the type of the add is integer, but for packed character
1838 // operations the type of the add needs to be char.
1839 void SuperWord::compute_vector_element_type() {
1840 #ifndef PRODUCT
1841   if (TraceSuperWord && Verbose)
1842     tty->print_cr("\ncompute_velt_type:");
1843 #endif
1844 
1845   // Initial type
1846   for (int i = 0; i < _block.length(); i++) {
1847     Node* n = _block.at(i);
1848     set_velt_type(n, container_type(n));
1849   }
1850 
1851   // Propagate integer narrowed type backwards through operations
1852   // that don't depend on higher order bits
1853   for (int i = _block.length() - 1; i >= 0; i--) {
1854     Node* n = _block.at(i);
1855     // Only integer types need be examined
1856     const Type* vtn = velt_type(n);
1857     if (vtn->basic_type() == T_INT) {
1858       uint start, end;
1859       VectorNode::vector_operands(n, &start, &end);
1860 
1861       for (uint j = start; j < end; j++) {
1862         Node* in  = n->in(j);
1863         // Don't propagate through a memory
1864         if (!in->is_Mem() && in_bb(in) && velt_type(in)->basic_type() == T_INT &&
1865             data_size(n) < data_size(in)) {
1866           bool same_type = true;
1867           for (DUIterator_Fast kmax, k = in->fast_outs(kmax); k < kmax; k++) {
1868             Node *use = in->fast_out(k);
1869             if (!in_bb(use) || !same_velt_type(use, n)) {
1870               same_type = false;
1871               break;
1872             }
1873           }
1874           if (same_type) {
1875             // For right shifts of small integer types (bool, byte, char, short)
1876             // we need precise information about sign-ness. Only Load nodes have
1877             // this information because Store nodes are the same for signed and
1878             // unsigned values. And any arithmetic operation after a load may
1879             // expand a value to signed Int so such right shifts can't be used
1880             // because vector elements do not have upper bits of Int.
1881             const Type* vt = vtn;
1882             if (VectorNode::is_shift(in)) {
1883               Node* load = in->in(1);
1884               if (load->is_Load() && in_bb(load) && (velt_type(load)->basic_type() == T_INT)) {
1885                 vt = velt_type(load);
1886               } else if (in->Opcode() != Op_LShiftI) {
1887                 // Widen type to Int to avoid creation of right shift vector
1888                 // (align + data_size(s1) check in stmts_can_pack() will fail).
1889                 // Note, left shifts work regardless type.
1890                 vt = TypeInt::INT;
1891               }
1892             }
1893             set_velt_type(in, vt);
1894           }
1895         }
1896       }
1897     }
1898   }
1899 #ifndef PRODUCT
1900   if (TraceSuperWord && Verbose) {
1901     for (int i = 0; i < _block.length(); i++) {
1902       Node* n = _block.at(i);
1903       velt_type(n)->dump();
1904       tty->print("\t");
1905       n->dump();
1906     }
1907   }
1908 #endif
1909 }
1910 
1911 //------------------------------memory_alignment---------------------------
1912 // Alignment within a vector memory reference
1913 int SuperWord::memory_alignment(MemNode* s, int iv_adjust) {
1914   SWPointer p(s, this);
1915   if (!p.valid()) {
1916     return bottom_align;
1917   }
1918   int vw = vector_width_in_bytes(s);
1919   if (vw < 2) {
1920     return bottom_align; // No vectors for this type
1921   }
1922   int offset  = p.offset_in_bytes();
1923   offset     += iv_adjust*p.memory_size();
1924   int off_rem = offset % vw;
1925   int off_mod = off_rem >= 0 ? off_rem : off_rem + vw;
1926   return off_mod;
1927 }
1928 
1929 //---------------------------container_type---------------------------
1930 // Smallest type containing range of values
1931 const Type* SuperWord::container_type(Node* n) {
1932   if (n->is_Mem()) {
1933     BasicType bt = n->as_Mem()->memory_type();
1934     if (n->is_Store() && (bt == T_CHAR)) {
1935       // Use T_SHORT type instead of T_CHAR for stored values because any
1936       // preceding arithmetic operation extends values to signed Int.
1937       bt = T_SHORT;
1938     }
1939     if (n->Opcode() == Op_LoadUB) {
1940       // Adjust type for unsigned byte loads, it is important for right shifts.
1941       // T_BOOLEAN is used because there is no basic type representing type
1942       // TypeInt::UBYTE. Use of T_BOOLEAN for vectors is fine because only
1943       // size (one byte) and sign is important.
1944       bt = T_BOOLEAN;
1945     }
1946     return Type::get_const_basic_type(bt);
1947   }
1948   const Type* t = _igvn.type(n);
1949   if (t->basic_type() == T_INT) {
1950     // A narrow type of arithmetic operations will be determined by
1951     // propagating the type of memory operations.
1952     return TypeInt::INT;
1953   }
1954   return t;
1955 }
1956 
1957 bool SuperWord::same_velt_type(Node* n1, Node* n2) {
1958   const Type* vt1 = velt_type(n1);
1959   const Type* vt2 = velt_type(n2);
1960   if (vt1->basic_type() == T_INT && vt2->basic_type() == T_INT) {
1961     // Compare vectors element sizes for integer types.
1962     return data_size(n1) == data_size(n2);
1963   }
1964   return vt1 == vt2;
1965 }
1966 
1967 //------------------------------in_packset---------------------------
1968 // Are s1 and s2 in a pack pair and ordered as s1,s2?
1969 bool SuperWord::in_packset(Node* s1, Node* s2) {
1970   for (int i = 0; i < _packset.length(); i++) {
1971     Node_List* p = _packset.at(i);
1972     assert(p->size() == 2, "must be");
1973     if (p->at(0) == s1 && p->at(p->size()-1) == s2) {
1974       return true;
1975     }
1976   }
1977   return false;
1978 }
1979 
1980 //------------------------------in_pack---------------------------
1981 // Is s in pack p?
1982 Node_List* SuperWord::in_pack(Node* s, Node_List* p) {
1983   for (uint i = 0; i < p->size(); i++) {
1984     if (p->at(i) == s) {
1985       return p;
1986     }
1987   }
1988   return NULL;
1989 }
1990 
1991 //------------------------------remove_pack_at---------------------------
1992 // Remove the pack at position pos in the packset
1993 void SuperWord::remove_pack_at(int pos) {
1994   Node_List* p = _packset.at(pos);
1995   for (uint i = 0; i < p->size(); i++) {
1996     Node* s = p->at(i);
1997     set_my_pack(s, NULL);
1998   }
1999   _packset.remove_at(pos);
2000 }
2001 
2002 //------------------------------executed_first---------------------------
2003 // Return the node executed first in pack p.  Uses the RPO block list
2004 // to determine order.
2005 Node* SuperWord::executed_first(Node_List* p) {
2006   Node* n = p->at(0);
2007   int n_rpo = bb_idx(n);
2008   for (uint i = 1; i < p->size(); i++) {
2009     Node* s = p->at(i);
2010     int s_rpo = bb_idx(s);
2011     if (s_rpo < n_rpo) {
2012       n = s;
2013       n_rpo = s_rpo;
2014     }
2015   }
2016   return n;
2017 }
2018 
2019 //------------------------------executed_last---------------------------
2020 // Return the node executed last in pack p.
2021 Node* SuperWord::executed_last(Node_List* p) {
2022   Node* n = p->at(0);
2023   int n_rpo = bb_idx(n);
2024   for (uint i = 1; i < p->size(); i++) {
2025     Node* s = p->at(i);
2026     int s_rpo = bb_idx(s);
2027     if (s_rpo > n_rpo) {
2028       n = s;
2029       n_rpo = s_rpo;
2030     }
2031   }
2032   return n;
2033 }
2034 
2035 LoadNode::ControlDependency SuperWord::control_dependency(Node_List* p) {
2036   LoadNode::ControlDependency dep = LoadNode::DependsOnlyOnTest;
2037   for (uint i = 0; i < p->size(); i++) {
2038     Node* n = p->at(i);
2039     assert(n->is_Load(), "only meaningful for loads");
2040     if (!n->depends_only_on_test()) {
2041       dep = LoadNode::Pinned;
2042     }
2043   }
2044   return dep;
2045 }
2046 
2047 
2048 //----------------------------align_initial_loop_index---------------------------
2049 // Adjust pre-loop limit so that in main loop, a load/store reference
2050 // to align_to_ref will be a position zero in the vector.
2051 //   (iv + k) mod vector_align == 0
2052 void SuperWord::align_initial_loop_index(MemNode* align_to_ref) {
2053   CountedLoopNode *main_head = lp()->as_CountedLoop();
2054   assert(main_head->is_main_loop(), "");
2055   CountedLoopEndNode* pre_end = get_pre_loop_end(main_head);
2056   assert(pre_end != NULL, "we must have a correct pre-loop");
2057   Node *pre_opaq1 = pre_end->limit();
2058   assert(pre_opaq1->Opcode() == Op_Opaque1, "");
2059   Opaque1Node *pre_opaq = (Opaque1Node*)pre_opaq1;
2060   Node *lim0 = pre_opaq->in(1);
2061 
2062   // Where we put new limit calculations
2063   Node *pre_ctrl = pre_end->loopnode()->in(LoopNode::EntryControl);
2064 
2065   // Ensure the original loop limit is available from the
2066   // pre-loop Opaque1 node.
2067   Node *orig_limit = pre_opaq->original_loop_limit();
2068   assert(orig_limit != NULL && _igvn.type(orig_limit) != Type::TOP, "");
2069 
2070   SWPointer align_to_ref_p(align_to_ref, this);
2071   assert(align_to_ref_p.valid(), "sanity");
2072 
2073   // Given:
2074   //     lim0 == original pre loop limit
2075   //     V == v_align (power of 2)
2076   //     invar == extra invariant piece of the address expression
2077   //     e == offset [ +/- invar ]
2078   //
2079   // When reassociating expressions involving '%' the basic rules are:
2080   //     (a - b) % k == 0   =>  a % k == b % k
2081   // and:
2082   //     (a + b) % k == 0   =>  a % k == (k - b) % k
2083   //
2084   // For stride > 0 && scale > 0,
2085   //   Derive the new pre-loop limit "lim" such that the two constraints:
2086   //     (1) lim = lim0 + N           (where N is some positive integer < V)
2087   //     (2) (e + lim) % V == 0
2088   //   are true.
2089   //
2090   //   Substituting (1) into (2),
2091   //     (e + lim0 + N) % V == 0
2092   //   solve for N:
2093   //     N = (V - (e + lim0)) % V
2094   //   substitute back into (1), so that new limit
2095   //     lim = lim0 + (V - (e + lim0)) % V
2096   //
2097   // For stride > 0 && scale < 0
2098   //   Constraints:
2099   //     lim = lim0 + N
2100   //     (e - lim) % V == 0
2101   //   Solving for lim:
2102   //     (e - lim0 - N) % V == 0
2103   //     N = (e - lim0) % V
2104   //     lim = lim0 + (e - lim0) % V
2105   //
2106   // For stride < 0 && scale > 0
2107   //   Constraints:
2108   //     lim = lim0 - N
2109   //     (e + lim) % V == 0
2110   //   Solving for lim:
2111   //     (e + lim0 - N) % V == 0
2112   //     N = (e + lim0) % V
2113   //     lim = lim0 - (e + lim0) % V
2114   //
2115   // For stride < 0 && scale < 0
2116   //   Constraints:
2117   //     lim = lim0 - N
2118   //     (e - lim) % V == 0
2119   //   Solving for lim:
2120   //     (e - lim0 + N) % V == 0
2121   //     N = (V - (e - lim0)) % V
2122   //     lim = lim0 - (V - (e - lim0)) % V
2123 
2124   int vw = vector_width_in_bytes(align_to_ref);
2125   int stride   = iv_stride();
2126   int scale    = align_to_ref_p.scale_in_bytes();
2127   int elt_size = align_to_ref_p.memory_size();
2128   int v_align  = vw / elt_size;
2129   assert(v_align > 1, "sanity");
2130   int offset   = align_to_ref_p.offset_in_bytes() / elt_size;
2131   Node *offsn  = _igvn.intcon(offset);
2132 
2133   Node *e = offsn;
2134   if (align_to_ref_p.invar() != NULL) {
2135     // incorporate any extra invariant piece producing (offset +/- invar) >>> log2(elt)
2136     Node* log2_elt = _igvn.intcon(exact_log2(elt_size));
2137     Node* aref     = new (_phase->C) URShiftINode(align_to_ref_p.invar(), log2_elt);
2138     _igvn.register_new_node_with_optimizer(aref);
2139     _phase->set_ctrl(aref, pre_ctrl);
2140     if (align_to_ref_p.negate_invar()) {
2141       e = new (_phase->C) SubINode(e, aref);
2142     } else {
2143       e = new (_phase->C) AddINode(e, aref);
2144     }
2145     _igvn.register_new_node_with_optimizer(e);
2146     _phase->set_ctrl(e, pre_ctrl);
2147   }
2148   if (vw > ObjectAlignmentInBytes) {
2149     // incorporate base e +/- base && Mask >>> log2(elt)
2150     Node* xbase = new(_phase->C) CastP2XNode(NULL, align_to_ref_p.base());
2151     _igvn.register_new_node_with_optimizer(xbase);
2152 #ifdef _LP64
2153     xbase  = new (_phase->C) ConvL2INode(xbase);
2154     _igvn.register_new_node_with_optimizer(xbase);
2155 #endif
2156     Node* mask = _igvn.intcon(vw-1);
2157     Node* masked_xbase  = new (_phase->C) AndINode(xbase, mask);
2158     _igvn.register_new_node_with_optimizer(masked_xbase);
2159     Node* log2_elt = _igvn.intcon(exact_log2(elt_size));
2160     Node* bref     = new (_phase->C) URShiftINode(masked_xbase, log2_elt);
2161     _igvn.register_new_node_with_optimizer(bref);
2162     _phase->set_ctrl(bref, pre_ctrl);
2163     e = new (_phase->C) AddINode(e, bref);
2164     _igvn.register_new_node_with_optimizer(e);
2165     _phase->set_ctrl(e, pre_ctrl);
2166   }
2167 
2168   // compute e +/- lim0
2169   if (scale < 0) {
2170     e = new (_phase->C) SubINode(e, lim0);
2171   } else {
2172     e = new (_phase->C) AddINode(e, lim0);
2173   }
2174   _igvn.register_new_node_with_optimizer(e);
2175   _phase->set_ctrl(e, pre_ctrl);
2176 
2177   if (stride * scale > 0) {
2178     // compute V - (e +/- lim0)
2179     Node* va  = _igvn.intcon(v_align);
2180     e = new (_phase->C) SubINode(va, e);
2181     _igvn.register_new_node_with_optimizer(e);
2182     _phase->set_ctrl(e, pre_ctrl);
2183   }
2184   // compute N = (exp) % V
2185   Node* va_msk = _igvn.intcon(v_align - 1);
2186   Node* N = new (_phase->C) AndINode(e, va_msk);
2187   _igvn.register_new_node_with_optimizer(N);
2188   _phase->set_ctrl(N, pre_ctrl);
2189 
2190   //   substitute back into (1), so that new limit
2191   //     lim = lim0 + N
2192   Node* lim;
2193   if (stride < 0) {
2194     lim = new (_phase->C) SubINode(lim0, N);
2195   } else {
2196     lim = new (_phase->C) AddINode(lim0, N);
2197   }
2198   _igvn.register_new_node_with_optimizer(lim);
2199   _phase->set_ctrl(lim, pre_ctrl);
2200   Node* constrained =
2201     (stride > 0) ? (Node*) new (_phase->C) MinINode(lim, orig_limit)
2202                  : (Node*) new (_phase->C) MaxINode(lim, orig_limit);
2203   _igvn.register_new_node_with_optimizer(constrained);
2204   _phase->set_ctrl(constrained, pre_ctrl);
2205   _igvn.hash_delete(pre_opaq);
2206   pre_opaq->set_req(1, constrained);
2207 }
2208 
2209 //----------------------------get_pre_loop_end---------------------------
2210 // Find pre loop end from main loop.  Returns null if none.
2211 CountedLoopEndNode* SuperWord::get_pre_loop_end(CountedLoopNode* cl) {
2212   Node* ctrl = cl->in(LoopNode::EntryControl);
2213   if (!ctrl->is_IfTrue() && !ctrl->is_IfFalse()) return NULL;
2214   Node* iffm = ctrl->in(0);
2215   if (!iffm->is_If()) return NULL;
2216   Node* bolzm = iffm->in(1);
2217   if (!bolzm->is_Bool()) return NULL;
2218   Node* cmpzm = bolzm->in(1);
2219   if (!cmpzm->is_Cmp()) return NULL;
2220   Node* opqzm = cmpzm->in(2);
2221   // Can not optimize a loop if zero-trip Opaque1 node is optimized
2222   // away and then another round of loop opts attempted.
2223   if (opqzm->Opcode() != Op_Opaque1) {
2224     return NULL;
2225   }
2226   Node* p_f = iffm->in(0);
2227   if (!p_f->is_IfFalse()) return NULL;
2228   if (!p_f->in(0)->is_CountedLoopEnd()) return NULL;
2229   CountedLoopEndNode* pre_end = p_f->in(0)->as_CountedLoopEnd();
2230   CountedLoopNode* loop_node = pre_end->loopnode();
2231   if (loop_node == NULL || !loop_node->is_pre_loop()) return NULL;
2232   return pre_end;
2233 }
2234 
2235 
2236 //------------------------------init---------------------------
2237 void SuperWord::init() {
2238   _dg.init();
2239   _packset.clear();
2240   _disjoint_ptrs.clear();
2241   _block.clear();
2242   _data_entry.clear();
2243   _mem_slice_head.clear();
2244   _mem_slice_tail.clear();
2245   _node_info.clear();
2246   _align_to_ref = NULL;
2247   _lpt = NULL;
2248   _lp = NULL;
2249   _bb = NULL;
2250   _iv = NULL;
2251 }
2252 
2253 //------------------------------print_packset---------------------------
2254 void SuperWord::print_packset() {
2255 #ifndef PRODUCT
2256   tty->print_cr("packset");
2257   for (int i = 0; i < _packset.length(); i++) {
2258     tty->print_cr("Pack: %d", i);
2259     Node_List* p = _packset.at(i);
2260     print_pack(p);
2261   }
2262 #endif
2263 }
2264 
2265 //------------------------------print_pack---------------------------
2266 void SuperWord::print_pack(Node_List* p) {
2267   for (uint i = 0; i < p->size(); i++) {
2268     print_stmt(p->at(i));
2269   }
2270 }
2271 
2272 //------------------------------print_bb---------------------------
2273 void SuperWord::print_bb() {
2274 #ifndef PRODUCT
2275   tty->print_cr("\nBlock");
2276   for (int i = 0; i < _block.length(); i++) {
2277     Node* n = _block.at(i);
2278     tty->print("%d ", i);
2279     if (n) {
2280       n->dump();
2281     }
2282   }
2283 #endif
2284 }
2285 
2286 //------------------------------print_stmt---------------------------
2287 void SuperWord::print_stmt(Node* s) {
2288 #ifndef PRODUCT
2289   tty->print(" align: %d \t", alignment(s));
2290   s->dump();
2291 #endif
2292 }
2293 
2294 //------------------------------blank---------------------------
2295 char* SuperWord::blank(uint depth) {
2296   static char blanks[101];
2297   assert(depth < 101, "too deep");
2298   for (uint i = 0; i < depth; i++) blanks[i] = ' ';
2299   blanks[depth] = '\0';
2300   return blanks;
2301 }
2302 
2303 
2304 //==============================SWPointer===========================
2305 
2306 //----------------------------SWPointer------------------------
2307 SWPointer::SWPointer(MemNode* mem, SuperWord* slp) :
2308   _mem(mem), _slp(slp),  _base(NULL),  _adr(NULL),
2309   _scale(0), _offset(0), _invar(NULL), _negate_invar(false) {
2310 
2311   Node* adr = mem->in(MemNode::Address);
2312   if (!adr->is_AddP()) {
2313     assert(!valid(), "too complex");
2314     return;
2315   }
2316   // Match AddP(base, AddP(ptr, k*iv [+ invariant]), constant)
2317   Node* base = adr->in(AddPNode::Base);
2318   // The base address should be loop invariant
2319   if (!invariant(base)) {
2320     assert(!valid(), "base address is loop variant");
2321     return;
2322   }
2323   //unsafe reference could not be aligned appropriately without runtime checking
2324   if (base == NULL || base->bottom_type() == Type::TOP) {
2325     assert(!valid(), "unsafe access");
2326     return;
2327   }
2328   for (int i = 0; i < 3; i++) {
2329     if (!scaled_iv_plus_offset(adr->in(AddPNode::Offset))) {
2330       assert(!valid(), "too complex");
2331       return;
2332     }
2333     adr = adr->in(AddPNode::Address);
2334     if (base == adr || !adr->is_AddP()) {
2335       break; // stop looking at addp's
2336     }
2337   }
2338   _base = base;
2339   _adr  = adr;
2340   assert(valid(), "Usable");
2341 }
2342 
2343 // Following is used to create a temporary object during
2344 // the pattern match of an address expression.
2345 SWPointer::SWPointer(SWPointer* p) :
2346   _mem(p->_mem), _slp(p->_slp),  _base(NULL),  _adr(NULL),
2347   _scale(0), _offset(0), _invar(NULL), _negate_invar(false) {}
2348 
2349 //------------------------scaled_iv_plus_offset--------------------
2350 // Match: k*iv + offset
2351 // where: k is a constant that maybe zero, and
2352 //        offset is (k2 [+/- invariant]) where k2 maybe zero and invariant is optional
2353 bool SWPointer::scaled_iv_plus_offset(Node* n) {
2354   if (scaled_iv(n)) {
2355     return true;
2356   }
2357   if (offset_plus_k(n)) {
2358     return true;
2359   }
2360   int opc = n->Opcode();
2361   if (opc == Op_AddI) {
2362     if (scaled_iv(n->in(1)) && offset_plus_k(n->in(2))) {
2363       return true;
2364     }
2365     if (scaled_iv(n->in(2)) && offset_plus_k(n->in(1))) {
2366       return true;
2367     }
2368   } else if (opc == Op_SubI) {
2369     if (scaled_iv(n->in(1)) && offset_plus_k(n->in(2), true)) {
2370       return true;
2371     }
2372     if (scaled_iv(n->in(2)) && offset_plus_k(n->in(1))) {
2373       _scale *= -1;
2374       return true;
2375     }
2376   }
2377   return false;
2378 }
2379 
2380 //----------------------------scaled_iv------------------------
2381 // Match: k*iv where k is a constant that's not zero
2382 bool SWPointer::scaled_iv(Node* n) {
2383   if (_scale != 0) {
2384     return false;  // already found a scale
2385   }
2386   if (n == iv()) {
2387     _scale = 1;
2388     return true;
2389   }
2390   int opc = n->Opcode();
2391   if (opc == Op_MulI) {
2392     if (n->in(1) == iv() && n->in(2)->is_Con()) {
2393       _scale = n->in(2)->get_int();
2394       return true;
2395     } else if (n->in(2) == iv() && n->in(1)->is_Con()) {
2396       _scale = n->in(1)->get_int();
2397       return true;
2398     }
2399   } else if (opc == Op_LShiftI) {
2400     if (n->in(1) == iv() && n->in(2)->is_Con()) {
2401       _scale = 1 << n->in(2)->get_int();
2402       return true;
2403     }
2404   } else if (opc == Op_ConvI2L) {
2405     if (n->in(1)->Opcode() == Op_CastII &&
2406         n->in(1)->as_CastII()->has_range_check()) {
2407       // Skip range check dependent CastII nodes
2408       n = n->in(1);
2409     }
2410     if (scaled_iv_plus_offset(n->in(1))) {
2411       return true;
2412     }
2413   } else if (opc == Op_LShiftL) {
2414     if (!has_iv() && _invar == NULL) {
2415       // Need to preserve the current _offset value, so
2416       // create a temporary object for this expression subtree.
2417       // Hacky, so should re-engineer the address pattern match.
2418       SWPointer tmp(this);
2419       if (tmp.scaled_iv_plus_offset(n->in(1))) {
2420         if (tmp._invar == NULL) {
2421           int mult = 1 << n->in(2)->get_int();
2422           _scale   = tmp._scale  * mult;
2423           _offset += tmp._offset * mult;
2424           return true;
2425         }
2426       }
2427     }
2428   }
2429   return false;
2430 }
2431 
2432 //----------------------------offset_plus_k------------------------
2433 // Match: offset is (k [+/- invariant])
2434 // where k maybe zero and invariant is optional, but not both.
2435 bool SWPointer::offset_plus_k(Node* n, bool negate) {
2436   int opc = n->Opcode();
2437   if (opc == Op_ConI) {
2438     _offset += negate ? -(n->get_int()) : n->get_int();
2439     return true;
2440   } else if (opc == Op_ConL) {
2441     // Okay if value fits into an int
2442     const TypeLong* t = n->find_long_type();
2443     if (t->higher_equal(TypeLong::INT)) {
2444       jlong loff = n->get_long();
2445       jint  off  = (jint)loff;
2446       _offset += negate ? -off : loff;
2447       return true;
2448     }
2449     return false;
2450   }
2451   if (_invar != NULL) return false; // already have an invariant
2452   if (opc == Op_AddI) {
2453     if (n->in(2)->is_Con() && invariant(n->in(1))) {
2454       _negate_invar = negate;
2455       _invar = n->in(1);
2456       _offset += negate ? -(n->in(2)->get_int()) : n->in(2)->get_int();
2457       return true;
2458     } else if (n->in(1)->is_Con() && invariant(n->in(2))) {
2459       _offset += negate ? -(n->in(1)->get_int()) : n->in(1)->get_int();
2460       _negate_invar = negate;
2461       _invar = n->in(2);
2462       return true;
2463     }
2464   }
2465   if (opc == Op_SubI) {
2466     if (n->in(2)->is_Con() && invariant(n->in(1))) {
2467       _negate_invar = negate;
2468       _invar = n->in(1);
2469       _offset += !negate ? -(n->in(2)->get_int()) : n->in(2)->get_int();
2470       return true;
2471     } else if (n->in(1)->is_Con() && invariant(n->in(2))) {
2472       _offset += negate ? -(n->in(1)->get_int()) : n->in(1)->get_int();
2473       _negate_invar = !negate;
2474       _invar = n->in(2);
2475       return true;
2476     }
2477   }
2478   if (invariant(n)) {
2479     _negate_invar = negate;
2480     _invar = n;
2481     return true;
2482   }
2483   return false;
2484 }
2485 
2486 //----------------------------print------------------------
2487 void SWPointer::print() {
2488 #ifndef PRODUCT
2489   tty->print("base: %d  adr: %d  scale: %d  offset: %d  invar: %c%d\n",
2490              _base != NULL ? _base->_idx : 0,
2491              _adr  != NULL ? _adr->_idx  : 0,
2492              _scale, _offset,
2493              _negate_invar?'-':'+',
2494              _invar != NULL ? _invar->_idx : 0);
2495 #endif
2496 }
2497 
2498 // ========================= OrderedPair =====================
2499 
2500 const OrderedPair OrderedPair::initial;
2501 
2502 // ========================= SWNodeInfo =====================
2503 
2504 const SWNodeInfo SWNodeInfo::initial;
2505 
2506 
2507 // ============================ DepGraph ===========================
2508 
2509 //------------------------------make_node---------------------------
2510 // Make a new dependence graph node for an ideal node.
2511 DepMem* DepGraph::make_node(Node* node) {
2512   DepMem* m = new (_arena) DepMem(node);
2513   if (node != NULL) {
2514     assert(_map.at_grow(node->_idx) == NULL, "one init only");
2515     _map.at_put_grow(node->_idx, m);
2516   }
2517   return m;
2518 }
2519 
2520 //------------------------------make_edge---------------------------
2521 // Make a new dependence graph edge from dpred -> dsucc
2522 DepEdge* DepGraph::make_edge(DepMem* dpred, DepMem* dsucc) {
2523   DepEdge* e = new (_arena) DepEdge(dpred, dsucc, dsucc->in_head(), dpred->out_head());
2524   dpred->set_out_head(e);
2525   dsucc->set_in_head(e);
2526   return e;
2527 }
2528 
2529 // ========================== DepMem ========================
2530 
2531 //------------------------------in_cnt---------------------------
2532 int DepMem::in_cnt() {
2533   int ct = 0;
2534   for (DepEdge* e = _in_head; e != NULL; e = e->next_in()) ct++;
2535   return ct;
2536 }
2537 
2538 //------------------------------out_cnt---------------------------
2539 int DepMem::out_cnt() {
2540   int ct = 0;
2541   for (DepEdge* e = _out_head; e != NULL; e = e->next_out()) ct++;
2542   return ct;
2543 }
2544 
2545 //------------------------------print-----------------------------
2546 void DepMem::print() {
2547 #ifndef PRODUCT
2548   tty->print("  DepNode %d (", _node->_idx);
2549   for (DepEdge* p = _in_head; p != NULL; p = p->next_in()) {
2550     Node* pred = p->pred()->node();
2551     tty->print(" %d", pred != NULL ? pred->_idx : 0);
2552   }
2553   tty->print(") [");
2554   for (DepEdge* s = _out_head; s != NULL; s = s->next_out()) {
2555     Node* succ = s->succ()->node();
2556     tty->print(" %d", succ != NULL ? succ->_idx : 0);
2557   }
2558   tty->print_cr(" ]");
2559 #endif
2560 }
2561 
2562 // =========================== DepEdge =========================
2563 
2564 //------------------------------DepPreds---------------------------
2565 void DepEdge::print() {
2566 #ifndef PRODUCT
2567   tty->print_cr("DepEdge: %d [ %d ]", _pred->node()->_idx, _succ->node()->_idx);
2568 #endif
2569 }
2570 
2571 // =========================== DepPreds =========================
2572 // Iterator over predecessor edges in the dependence graph.
2573 
2574 //------------------------------DepPreds---------------------------
2575 DepPreds::DepPreds(Node* n, DepGraph& dg) {
2576   _n = n;
2577   _done = false;
2578   if (_n->is_Store() || _n->is_Load()) {
2579     _next_idx = MemNode::Address;
2580     _end_idx  = n->req();
2581     _dep_next = dg.dep(_n)->in_head();
2582   } else if (_n->is_Mem()) {
2583     _next_idx = 0;
2584     _end_idx  = 0;
2585     _dep_next = dg.dep(_n)->in_head();
2586   } else {
2587     _next_idx = 1;
2588     _end_idx  = _n->req();
2589     _dep_next = NULL;
2590   }
2591   next();
2592 }
2593 
2594 //------------------------------next---------------------------
2595 void DepPreds::next() {
2596   if (_dep_next != NULL) {
2597     _current  = _dep_next->pred()->node();
2598     _dep_next = _dep_next->next_in();
2599   } else if (_next_idx < _end_idx) {
2600     _current  = _n->in(_next_idx++);
2601   } else {
2602     _done = true;
2603   }
2604 }
2605 
2606 // =========================== DepSuccs =========================
2607 // Iterator over successor edges in the dependence graph.
2608 
2609 //------------------------------DepSuccs---------------------------
2610 DepSuccs::DepSuccs(Node* n, DepGraph& dg) {
2611   _n = n;
2612   _done = false;
2613   if (_n->is_Load()) {
2614     _next_idx = 0;
2615     _end_idx  = _n->outcnt();
2616     _dep_next = dg.dep(_n)->out_head();
2617   } else if (_n->is_Mem() || _n->is_Phi() && _n->bottom_type() == Type::MEMORY) {
2618     _next_idx = 0;
2619     _end_idx  = 0;
2620     _dep_next = dg.dep(_n)->out_head();
2621   } else {
2622     _next_idx = 0;
2623     _end_idx  = _n->outcnt();
2624     _dep_next = NULL;
2625   }
2626   next();
2627 }
2628 
2629 //-------------------------------next---------------------------
2630 void DepSuccs::next() {
2631   if (_dep_next != NULL) {
2632     _current  = _dep_next->succ()->node();
2633     _dep_next = _dep_next->next_out();
2634   } else if (_next_idx < _end_idx) {
2635     _current  = _n->raw_out(_next_idx++);
2636   } else {
2637     _done = true;
2638   }
2639 }