1 /*
   2  * Copyright (c) 2007, 2015, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  */
  23 
  24 #include "precompiled.hpp"
  25 #include "compiler/compileLog.hpp"
  26 #include "libadt/vectset.hpp"
  27 #include "memory/allocation.inline.hpp"
  28 #include "opto/addnode.hpp"
  29 #include "opto/callnode.hpp"
  30 #include "opto/castnode.hpp"
  31 #include "opto/convertnode.hpp"
  32 #include "opto/divnode.hpp"
  33 #include "opto/matcher.hpp"
  34 #include "opto/memnode.hpp"
  35 #include "opto/mulnode.hpp"
  36 #include "opto/opcodes.hpp"
  37 #include "opto/opaquenode.hpp"
  38 #include "opto/superword.hpp"
  39 #include "opto/vectornode.hpp"
  40 
  41 //
  42 //                  S U P E R W O R D   T R A N S F O R M
  43 //=============================================================================
  44 
  45 //------------------------------SuperWord---------------------------
  46 SuperWord::SuperWord(PhaseIdealLoop* phase) :
  47   _phase(phase),
  48   _igvn(phase->_igvn),
  49   _arena(phase->C->comp_arena()),
  50   _packset(arena(), 8,  0, NULL),         // packs for the current block
  51   _bb_idx(arena(), (int)(1.10 * phase->C->unique()), 0, 0), // node idx to index in bb
  52   _block(arena(), 8,  0, NULL),           // nodes in current block
  53   _data_entry(arena(), 8,  0, NULL),      // nodes with all inputs from outside
  54   _mem_slice_head(arena(), 8,  0, NULL),  // memory slice heads
  55   _mem_slice_tail(arena(), 8,  0, NULL),  // memory slice tails
  56   _node_info(arena(), 8,  0, SWNodeInfo::initial), // info needed per node
  57   _clone_map(phase->C->clone_map()),      // map of nodes created in cloning
  58   _align_to_ref(NULL),                    // memory reference to align vectors to
  59   _disjoint_ptrs(arena(), 8,  0, OrderedPair::initial), // runtime disambiguated pointer pairs
  60   _dg(_arena),                            // dependence graph
  61   _visited(arena()),                      // visited node set
  62   _post_visited(arena()),                 // post visited node set
  63   _n_idx_list(arena(), 8),                // scratch list of (node,index) pairs
  64   _stk(arena(), 8, 0, NULL),              // scratch stack of nodes
  65   _nlist(arena(), 8, 0, NULL),            // scratch list of nodes
  66   _lpt(NULL),                             // loop tree node
  67   _lp(NULL),                              // LoopNode
  68   _bb(NULL),                              // basic block
  69   _iv(NULL),                              // induction var
  70   _race_possible(false),                  // cases where SDMU is true
  71   _early_return(true),                    // analysis evaluations routine
  72   _num_work_vecs(0),                      // amount of vector work we have
  73   _num_reductions(0),                     // amount of reduction work we have
  74   _do_vector_loop(phase->C->do_vector_loop()),  // whether to do vectorization/simd style
  75   _ii_first(-1),                          // first loop generation index - only if do_vector_loop()
  76   _ii_last(-1),                           // last loop generation index - only if do_vector_loop()
  77   _ii_order(arena(), 8, 0, 0),
  78   _vector_loop_debug(phase->C->has_method() && phase->C->method_has_option("VectorizeDebug"))
  79 {}
  80 
  81 //------------------------------transform_loop---------------------------
  82 void SuperWord::transform_loop(IdealLoopTree* lpt, bool do_optimization) {
  83   assert(UseSuperWord, "should be");
  84   // Do vectors exist on this architecture?
  85   if (Matcher::vector_width_in_bytes(T_BYTE) < 2) return;
  86 
  87   assert(lpt->_head->is_CountedLoop(), "must be");
  88   CountedLoopNode *cl = lpt->_head->as_CountedLoop();
  89 
  90   if (!cl->is_valid_counted_loop()) return; // skip malformed counted loop
  91 
  92   if (!cl->is_main_loop() ) return; // skip normal, pre, and post loops
  93 
  94   // Check for no control flow in body (other than exit)
  95   Node *cl_exit = cl->loopexit();
  96   if (cl_exit->in(0) != lpt->_head) return;
  97 
  98   // Make sure the are no extra control users of the loop backedge
  99   if (cl->back_control()->outcnt() != 1) {
 100     return;
 101   }
 102 
 103   // Check for pre-loop ending with CountedLoopEnd(Bool(Cmp(x,Opaque1(limit))))
 104   CountedLoopEndNode* pre_end = get_pre_loop_end(cl);
 105   if (pre_end == NULL) return;
 106   Node *pre_opaq1 = pre_end->limit();
 107   if (pre_opaq1->Opcode() != Op_Opaque1) return;
 108 
 109   init(); // initialize data structures
 110 
 111   set_lpt(lpt);
 112   set_lp(cl);
 113 
 114   // For now, define one block which is the entire loop body
 115   set_bb(cl);
 116 
 117   if (do_optimization) {
 118     assert(_packset.length() == 0, "packset must be empty");
 119     SLP_extract();
 120   }
 121 }
 122 
 123 //------------------------------early unrolling analysis------------------------------
 124 void SuperWord::unrolling_analysis(CountedLoopNode *cl, int &local_loop_unroll_factor) {
 125   bool is_slp = true;
 126   ResourceMark rm;
 127   size_t ignored_size = lpt()->_body.size();
 128   int *ignored_loop_nodes = NEW_RESOURCE_ARRAY(int, ignored_size);
 129   Node_Stack nstack((int)ignored_size);
 130   Node *cl_exit = cl->loopexit();
 131 
 132   // First clear the entries
 133   for (uint i = 0; i < lpt()->_body.size(); i++) {
 134     ignored_loop_nodes[i] = -1;
 135   }
 136 
 137   int max_vector = Matcher::max_vector_size(T_INT);
 138 
 139   // Process the loop, some/all of the stack entries will not be in order, ergo
 140   // need to preprocess the ignored initial state before we process the loop
 141   for (uint i = 0; i < lpt()->_body.size(); i++) {
 142     Node* n = lpt()->_body.at(i);
 143     if (n == cl->incr() ||
 144       n->is_reduction() ||
 145       n->is_AddP() ||
 146       n->is_Cmp() ||
 147       n->is_IfTrue() ||
 148       n->is_CountedLoop() ||
 149       (n == cl_exit)) {
 150       ignored_loop_nodes[i] = n->_idx;
 151       continue;
 152     }
 153 
 154     if (n->is_If()) {
 155       IfNode *iff = n->as_If();
 156       if (iff->_fcnt != COUNT_UNKNOWN && iff->_prob != PROB_UNKNOWN) {
 157         if (lpt()->is_loop_exit(iff)) {
 158           ignored_loop_nodes[i] = n->_idx;
 159           continue;
 160         }
 161       }
 162     }
 163 
 164     if (n->is_Phi() && (n->bottom_type() == Type::MEMORY)) {
 165       Node* n_tail = n->in(LoopNode::LoopBackControl);
 166       if (n_tail != n->in(LoopNode::EntryControl)) {
 167         if (!n_tail->is_Mem()) {
 168           is_slp = false;
 169           break;
 170         }
 171       }
 172     }
 173 
 174     // This must happen after check of phi/if
 175     if (n->is_Phi() || n->is_If()) {
 176       ignored_loop_nodes[i] = n->_idx;
 177       continue;
 178     }
 179 
 180     if (n->is_LoadStore() || n->is_MergeMem() ||
 181       (n->is_Proj() && !n->as_Proj()->is_CFG())) {
 182       is_slp = false;
 183       break;
 184     }
 185 
 186     // Ignore nodes with non-primitive type.
 187     BasicType bt;
 188     if (n->is_Mem()) {
 189       bt = n->as_Mem()->memory_type();
 190     } else {
 191       bt = n->bottom_type()->basic_type();
 192     }
 193     if (is_java_primitive(bt) == false) {
 194       ignored_loop_nodes[i] = n->_idx;
 195       continue;
 196     }
 197 
 198     if (n->is_Mem()) {
 199       MemNode* current = n->as_Mem();
 200       Node* adr = n->in(MemNode::Address);
 201       Node* n_ctrl = _phase->get_ctrl(adr);
 202 
 203       // save a queue of post process nodes
 204       if (n_ctrl != NULL && lpt()->is_member(_phase->get_loop(n_ctrl))) {
 205         // Process the memory expression
 206         int stack_idx = 0;
 207         bool have_side_effects = true;
 208         if (adr->is_AddP() == false) {
 209           nstack.push(adr, stack_idx++);
 210         } else {
 211           // Mark the components of the memory operation in nstack
 212           SWPointer p1(current, this, &nstack, true);
 213           have_side_effects = p1.node_stack()->is_nonempty();
 214         }
 215 
 216         // Process the pointer stack
 217         while (have_side_effects) {
 218           Node* pointer_node = nstack.node();
 219           for (uint j = 0; j < lpt()->_body.size(); j++) {
 220             Node* cur_node = lpt()->_body.at(j);
 221             if (cur_node == pointer_node) {
 222               ignored_loop_nodes[j] = cur_node->_idx;
 223               break;
 224             }
 225           }
 226           nstack.pop();
 227           have_side_effects = nstack.is_nonempty();
 228         }
 229       }
 230     }
 231   }
 232 
 233   if (is_slp) {
 234     // Now we try to find the maximum supported consistent vector which the machine
 235     // description can use
 236     for (uint i = 0; i < lpt()->_body.size(); i++) {
 237       if (ignored_loop_nodes[i] != -1) continue;
 238 
 239       BasicType bt;
 240       Node* n = lpt()->_body.at(i);
 241       if (n->is_Mem()) {
 242         bt = n->as_Mem()->memory_type();
 243       } else {
 244         bt = n->bottom_type()->basic_type();
 245       }
 246       if (is_java_primitive(bt) == false) continue;
 247 
 248       int cur_max_vector = Matcher::max_vector_size(bt);
 249 
 250       // If a max vector exists which is not larger than _local_loop_unroll_factor
 251       // stop looking, we already have the max vector to map to.
 252       if (cur_max_vector <= local_loop_unroll_factor) {
 253         is_slp = false;
 254 #ifndef PRODUCT
 255         if (TraceSuperWordLoopUnrollAnalysis) {
 256           tty->print_cr("slp analysis fails: unroll limit equals max vector\n");
 257         }
 258 #endif
 259         break;
 260       }
 261 
 262       // Map the maximal common vector
 263       if (VectorNode::implemented(n->Opcode(), cur_max_vector, bt)) {
 264         if (cur_max_vector < max_vector) {
 265           max_vector = cur_max_vector;
 266         }
 267       }
 268     }
 269     if (is_slp) {
 270       local_loop_unroll_factor = max_vector;
 271     }
 272     cl->mark_passed_slp();
 273     cl->set_slp_max_unroll(local_loop_unroll_factor);
 274   }
 275 }
 276 
 277 //------------------------------SLP_extract---------------------------
 278 // Extract the superword level parallelism
 279 //
 280 // 1) A reverse post-order of nodes in the block is constructed.  By scanning
 281 //    this list from first to last, all definitions are visited before their uses.
 282 //
 283 // 2) A point-to-point dependence graph is constructed between memory references.
 284 //    This simplies the upcoming "independence" checker.
 285 //
 286 // 3) The maximum depth in the node graph from the beginning of the block
 287 //    to each node is computed.  This is used to prune the graph search
 288 //    in the independence checker.
 289 //
 290 // 4) For integer types, the necessary bit width is propagated backwards
 291 //    from stores to allow packed operations on byte, char, and short
 292 //    integers.  This reverses the promotion to type "int" that javac
 293 //    did for operations like: char c1,c2,c3;  c1 = c2 + c3.
 294 //
 295 // 5) One of the memory references is picked to be an aligned vector reference.
 296 //    The pre-loop trip count is adjusted to align this reference in the
 297 //    unrolled body.
 298 //
 299 // 6) The initial set of pack pairs is seeded with memory references.
 300 //
 301 // 7) The set of pack pairs is extended by following use->def and def->use links.
 302 //
 303 // 8) The pairs are combined into vector sized packs.
 304 //
 305 // 9) Reorder the memory slices to co-locate members of the memory packs.
 306 //
 307 // 10) Generate ideal vector nodes for the final set of packs and where necessary,
 308 //    inserting scalar promotion, vector creation from multiple scalars, and
 309 //    extraction of scalar values from vectors.
 310 //
 311 void SuperWord::SLP_extract() {
 312 
 313 #ifndef PRODUCT
 314   if (_do_vector_loop && TraceSuperWord) {
 315     tty->print("SuperWord::SLP_extract\n");
 316     tty->print("input loop\n");
 317     _lpt->dump_head();
 318     _lpt->dump();
 319     for (uint i = 0; i < _lpt->_body.size(); i++) {
 320       _lpt->_body.at(i)->dump();
 321     }
 322   }
 323 #endif
 324   // Ready the block
 325   if (!construct_bb()) {
 326     return; // Exit if no interesting nodes or complex graph.
 327   }
 328   // build    _dg, _disjoint_ptrs
 329   dependence_graph();
 330 
 331   // compute function depth(Node*)
 332   compute_max_depth();
 333 
 334   if (_do_vector_loop) {
 335     if (mark_generations() != -1) {
 336       hoist_loads_in_graph(); // this only rebuild the graph; all basic structs need rebuild explicitly
 337 
 338       if (!construct_bb()) {
 339         return; // Exit if no interesting nodes or complex graph.
 340       }
 341       dependence_graph();
 342       compute_max_depth();
 343     }
 344 
 345 #ifndef PRODUCT
 346     if (TraceSuperWord) {
 347       tty->print_cr("\nSuperWord::_do_vector_loop: graph after hoist_loads_in_graph");
 348       _lpt->dump_head();
 349       for (int j = 0; j < _block.length(); j++) {
 350         Node* n = _block.at(j);
 351         int d = depth(n);
 352         for (int i = 0;  i < d; i++) tty->print("%s", "  ");
 353         tty->print("%d :", d);
 354         n->dump();
 355       }
 356     }
 357 #endif
 358   }
 359 
 360   compute_vector_element_type();
 361 
 362   // Attempt vectorization
 363 
 364   find_adjacent_refs();
 365 
 366   extend_packlist();
 367 
 368   if (_do_vector_loop) {
 369     if (_packset.length() == 0) {
 370 #ifndef PRODUCT
 371       if (TraceSuperWord) {
 372         tty->print_cr("\nSuperWord::_do_vector_loop DFA could not build packset, now trying to build anyway");
 373       }
 374 #endif
 375       pack_parallel();
 376     }
 377   }
 378 
 379   combine_packs();
 380 
 381   construct_my_pack_map();
 382 
 383   filter_packs();
 384 
 385   schedule();
 386 
 387   output();
 388 }
 389 
 390 //------------------------------find_adjacent_refs---------------------------
 391 // Find the adjacent memory references and create pack pairs for them.
 392 // This is the initial set of packs that will then be extended by
 393 // following use->def and def->use links.  The align positions are
 394 // assigned relative to the reference "align_to_ref"
 395 void SuperWord::find_adjacent_refs() {
 396   // Get list of memory operations
 397   Node_List memops;
 398   for (int i = 0; i < _block.length(); i++) {
 399     Node* n = _block.at(i);
 400     if (n->is_Mem() && !n->is_LoadStore() && in_bb(n) &&
 401         is_java_primitive(n->as_Mem()->memory_type())) {
 402       int align = memory_alignment(n->as_Mem(), 0);
 403       if (align != bottom_align) {
 404         memops.push(n);
 405       }
 406     }
 407   }
 408 
 409   Node_List align_to_refs;
 410   int best_iv_adjustment = 0;
 411   MemNode* best_align_to_mem_ref = NULL;
 412 
 413   while (memops.size() != 0) {
 414     // Find a memory reference to align to.
 415     MemNode* mem_ref = find_align_to_ref(memops);
 416     if (mem_ref == NULL) break;
 417     align_to_refs.push(mem_ref);
 418     int iv_adjustment = get_iv_adjustment(mem_ref);
 419 
 420     if (best_align_to_mem_ref == NULL) {
 421       // Set memory reference which is the best from all memory operations
 422       // to be used for alignment. The pre-loop trip count is modified to align
 423       // this reference to a vector-aligned address.
 424       best_align_to_mem_ref = mem_ref;
 425       best_iv_adjustment = iv_adjustment;
 426     }
 427 
 428     SWPointer align_to_ref_p(mem_ref, this, NULL, false);
 429     // Set alignment relative to "align_to_ref" for all related memory operations.
 430     for (int i = memops.size() - 1; i >= 0; i--) {
 431       MemNode* s = memops.at(i)->as_Mem();
 432       if (isomorphic(s, mem_ref)) {
 433         SWPointer p2(s, this, NULL, false);
 434         if (p2.comparable(align_to_ref_p)) {
 435           int align = memory_alignment(s, iv_adjustment);
 436           set_alignment(s, align);
 437         }
 438       }
 439     }
 440 
 441     // Create initial pack pairs of memory operations for which
 442     // alignment is set and vectors will be aligned.
 443     bool create_pack = true;
 444     if (memory_alignment(mem_ref, best_iv_adjustment) == 0 || _do_vector_loop) {
 445       if (!Matcher::misaligned_vectors_ok()) {
 446         int vw = vector_width(mem_ref);
 447         int vw_best = vector_width(best_align_to_mem_ref);
 448         if (vw > vw_best) {
 449           // Do not vectorize a memory access with more elements per vector
 450           // if unaligned memory access is not allowed because number of
 451           // iterations in pre-loop will be not enough to align it.
 452           create_pack = false;
 453         } else {
 454           SWPointer p2(best_align_to_mem_ref, this, NULL, false);
 455           if (align_to_ref_p.invar() != p2.invar()) {
 456             // Do not vectorize memory accesses with different invariants
 457             // if unaligned memory accesses are not allowed.
 458             create_pack = false;
 459           }
 460         }
 461       }
 462     } else {
 463       if (same_velt_type(mem_ref, best_align_to_mem_ref)) {
 464         // Can't allow vectorization of unaligned memory accesses with the
 465         // same type since it could be overlapped accesses to the same array.
 466         create_pack = false;
 467       } else {
 468         // Allow independent (different type) unaligned memory operations
 469         // if HW supports them.
 470         if (!Matcher::misaligned_vectors_ok()) {
 471           create_pack = false;
 472         } else {
 473           // Check if packs of the same memory type but
 474           // with a different alignment were created before.
 475           for (uint i = 0; i < align_to_refs.size(); i++) {
 476             MemNode* mr = align_to_refs.at(i)->as_Mem();
 477             if (same_velt_type(mr, mem_ref) &&
 478                 memory_alignment(mr, iv_adjustment) != 0)
 479               create_pack = false;
 480           }
 481         }
 482       }
 483     }
 484     if (create_pack) {
 485       for (uint i = 0; i < memops.size(); i++) {
 486         Node* s1 = memops.at(i);
 487         int align = alignment(s1);
 488         if (align == top_align) continue;
 489         for (uint j = 0; j < memops.size(); j++) {
 490           Node* s2 = memops.at(j);
 491           if (alignment(s2) == top_align) continue;
 492           if (s1 != s2 && are_adjacent_refs(s1, s2)) {
 493             if (stmts_can_pack(s1, s2, align)) {
 494               Node_List* pair = new Node_List();
 495               pair->push(s1);
 496               pair->push(s2);
 497               if (!_do_vector_loop || _clone_map.idx(s1->_idx) == _clone_map.idx(s2->_idx)) {
 498                 _packset.append(pair);
 499               }
 500             }
 501           }
 502         }
 503       }
 504     } else { // Don't create unaligned pack
 505       // First, remove remaining memory ops of the same type from the list.
 506       for (int i = memops.size() - 1; i >= 0; i--) {
 507         MemNode* s = memops.at(i)->as_Mem();
 508         if (same_velt_type(s, mem_ref)) {
 509           memops.remove(i);
 510         }
 511       }
 512 
 513       // Second, remove already constructed packs of the same type.
 514       for (int i = _packset.length() - 1; i >= 0; i--) {
 515         Node_List* p = _packset.at(i);
 516         MemNode* s = p->at(0)->as_Mem();
 517         if (same_velt_type(s, mem_ref)) {
 518           remove_pack_at(i);
 519         }
 520       }
 521 
 522       // If needed find the best memory reference for loop alignment again.
 523       if (same_velt_type(mem_ref, best_align_to_mem_ref)) {
 524         // Put memory ops from remaining packs back on memops list for
 525         // the best alignment search.
 526         uint orig_msize = memops.size();
 527         for (int i = 0; i < _packset.length(); i++) {
 528           Node_List* p = _packset.at(i);
 529           MemNode* s = p->at(0)->as_Mem();
 530           assert(!same_velt_type(s, mem_ref), "sanity");
 531           memops.push(s);
 532         }
 533         MemNode* best_align_to_mem_ref = find_align_to_ref(memops);
 534         if (best_align_to_mem_ref == NULL) break;
 535         best_iv_adjustment = get_iv_adjustment(best_align_to_mem_ref);
 536         // Restore list.
 537         while (memops.size() > orig_msize)
 538           (void)memops.pop();
 539       }
 540     } // unaligned memory accesses
 541 
 542     // Remove used mem nodes.
 543     for (int i = memops.size() - 1; i >= 0; i--) {
 544       MemNode* m = memops.at(i)->as_Mem();
 545       if (alignment(m) != top_align) {
 546         memops.remove(i);
 547       }
 548     }
 549 
 550   } // while (memops.size() != 0
 551   set_align_to_ref(best_align_to_mem_ref);
 552 
 553 #ifndef PRODUCT
 554   if (TraceSuperWord) {
 555     tty->print_cr("\nAfter find_adjacent_refs");
 556     print_packset();
 557   }
 558 #endif
 559 }
 560 
 561 //------------------------------find_align_to_ref---------------------------
 562 // Find a memory reference to align the loop induction variable to.
 563 // Looks first at stores then at loads, looking for a memory reference
 564 // with the largest number of references similar to it.
 565 MemNode* SuperWord::find_align_to_ref(Node_List &memops) {
 566   GrowableArray<int> cmp_ct(arena(), memops.size(), memops.size(), 0);
 567 
 568   // Count number of comparable memory ops
 569   for (uint i = 0; i < memops.size(); i++) {
 570     MemNode* s1 = memops.at(i)->as_Mem();
 571     SWPointer p1(s1, this, NULL, false);
 572     // Discard if pre loop can't align this reference
 573     if (!ref_is_alignable(p1)) {
 574       *cmp_ct.adr_at(i) = 0;
 575       continue;
 576     }
 577     for (uint j = i+1; j < memops.size(); j++) {
 578       MemNode* s2 = memops.at(j)->as_Mem();
 579       if (isomorphic(s1, s2)) {
 580         SWPointer p2(s2, this, NULL, false);
 581         if (p1.comparable(p2)) {
 582           (*cmp_ct.adr_at(i))++;
 583           (*cmp_ct.adr_at(j))++;
 584         }
 585       }
 586     }
 587   }
 588 
 589   // Find Store (or Load) with the greatest number of "comparable" references,
 590   // biggest vector size, smallest data size and smallest iv offset.
 591   int max_ct        = 0;
 592   int max_vw        = 0;
 593   int max_idx       = -1;
 594   int min_size      = max_jint;
 595   int min_iv_offset = max_jint;
 596   for (uint j = 0; j < memops.size(); j++) {
 597     MemNode* s = memops.at(j)->as_Mem();
 598     if (s->is_Store()) {
 599       int vw = vector_width_in_bytes(s);
 600       assert(vw > 1, "sanity");
 601       SWPointer p(s, this, NULL, false);
 602       if (cmp_ct.at(j) >  max_ct ||
 603           cmp_ct.at(j) == max_ct &&
 604             (vw >  max_vw ||
 605              vw == max_vw &&
 606               (data_size(s) <  min_size ||
 607                data_size(s) == min_size &&
 608                  (p.offset_in_bytes() < min_iv_offset)))) {
 609         max_ct = cmp_ct.at(j);
 610         max_vw = vw;
 611         max_idx = j;
 612         min_size = data_size(s);
 613         min_iv_offset = p.offset_in_bytes();
 614       }
 615     }
 616   }
 617   // If no stores, look at loads
 618   if (max_ct == 0) {
 619     for (uint j = 0; j < memops.size(); j++) {
 620       MemNode* s = memops.at(j)->as_Mem();
 621       if (s->is_Load()) {
 622         int vw = vector_width_in_bytes(s);
 623         assert(vw > 1, "sanity");
 624         SWPointer p(s, this, NULL, false);
 625         if (cmp_ct.at(j) >  max_ct ||
 626             cmp_ct.at(j) == max_ct &&
 627               (vw >  max_vw ||
 628                vw == max_vw &&
 629                 (data_size(s) <  min_size ||
 630                  data_size(s) == min_size &&
 631                    (p.offset_in_bytes() < min_iv_offset)))) {
 632           max_ct = cmp_ct.at(j);
 633           max_vw = vw;
 634           max_idx = j;
 635           min_size = data_size(s);
 636           min_iv_offset = p.offset_in_bytes();
 637         }
 638       }
 639     }
 640   }
 641 
 642 #ifdef ASSERT
 643   if (TraceSuperWord && Verbose) {
 644     tty->print_cr("\nVector memops after find_align_to_ref");
 645     for (uint i = 0; i < memops.size(); i++) {
 646       MemNode* s = memops.at(i)->as_Mem();
 647       s->dump();
 648     }
 649   }
 650 #endif
 651 
 652   if (max_ct > 0) {
 653 #ifdef ASSERT
 654     if (TraceSuperWord) {
 655       tty->print("\nVector align to node: ");
 656       memops.at(max_idx)->as_Mem()->dump();
 657     }
 658 #endif
 659     return memops.at(max_idx)->as_Mem();
 660   }
 661   return NULL;
 662 }
 663 
 664 //------------------------------ref_is_alignable---------------------------
 665 // Can the preloop align the reference to position zero in the vector?
 666 bool SuperWord::ref_is_alignable(SWPointer& p) {
 667   if (!p.has_iv()) {
 668     return true;   // no induction variable
 669   }
 670   CountedLoopEndNode* pre_end = get_pre_loop_end(lp()->as_CountedLoop());
 671   assert(pre_end != NULL, "we must have a correct pre-loop");
 672   assert(pre_end->stride_is_con(), "pre loop stride is constant");
 673   int preloop_stride = pre_end->stride_con();
 674 
 675   int span = preloop_stride * p.scale_in_bytes();
 676   int mem_size = p.memory_size();
 677   int offset   = p.offset_in_bytes();
 678   // Stride one accesses are alignable if offset is aligned to memory operation size.
 679   // Offset can be unaligned when UseUnalignedAccesses is used.
 680   if (ABS(span) == mem_size && (ABS(offset) % mem_size) == 0) {
 681     return true;
 682   }
 683   // If the initial offset from start of the object is computable,
 684   // check if the pre-loop can align the final offset accordingly.
 685   //
 686   // In other words: Can we find an i such that the offset
 687   // after i pre-loop iterations is aligned to vw?
 688   //   (init_offset + pre_loop) % vw == 0              (1)
 689   // where
 690   //   pre_loop = i * span
 691   // is the number of bytes added to the offset by i pre-loop iterations.
 692   //
 693   // For this to hold we need pre_loop to increase init_offset by
 694   //   pre_loop = vw - (init_offset % vw)
 695   //
 696   // This is only possible if pre_loop is divisible by span because each
 697   // pre-loop iteration increases the initial offset by 'span' bytes:
 698   //   (vw - (init_offset % vw)) % span == 0
 699   //
 700   int vw = vector_width_in_bytes(p.mem());
 701   assert(vw > 1, "sanity");
 702   Node* init_nd = pre_end->init_trip();
 703   if (init_nd->is_Con() && p.invar() == NULL) {
 704     int init = init_nd->bottom_type()->is_int()->get_con();
 705     int init_offset = init * p.scale_in_bytes() + offset;
 706     assert(init_offset >= 0, "positive offset from object start");
 707     if (vw % span == 0) {
 708       // If vm is a multiple of span, we use formula (1).
 709       if (span > 0) {
 710         return (vw - (init_offset % vw)) % span == 0;
 711       } else {
 712         assert(span < 0, "nonzero stride * scale");
 713         return (init_offset % vw) % -span == 0;
 714       }
 715     } else if (span % vw == 0) {
 716       // If span is a multiple of vw, we can simplify formula (1) to:
 717       //   (init_offset + i * span) % vw == 0
 718       //     =>
 719       //   (init_offset % vw) + ((i * span) % vw) == 0
 720       //     =>
 721       //   init_offset % vw == 0
 722       //
 723       // Because we add a multiple of vw to the initial offset, the final
 724       // offset is a multiple of vw if and only if init_offset is a multiple.
 725       //
 726       return (init_offset % vw) == 0;
 727     }
 728   }
 729   return false;
 730 }
 731 
 732 //---------------------------get_iv_adjustment---------------------------
 733 // Calculate loop's iv adjustment for this memory ops.
 734 int SuperWord::get_iv_adjustment(MemNode* mem_ref) {
 735   SWPointer align_to_ref_p(mem_ref, this, NULL, false);
 736   int offset = align_to_ref_p.offset_in_bytes();
 737   int scale  = align_to_ref_p.scale_in_bytes();
 738   int elt_size = align_to_ref_p.memory_size();
 739   int vw       = vector_width_in_bytes(mem_ref);
 740   assert(vw > 1, "sanity");
 741   int iv_adjustment;
 742   if (scale != 0) {
 743     int stride_sign = (scale * iv_stride()) > 0 ? 1 : -1;
 744     // At least one iteration is executed in pre-loop by default. As result
 745     // several iterations are needed to align memory operations in main-loop even
 746     // if offset is 0.
 747     int iv_adjustment_in_bytes = (stride_sign * vw - (offset % vw));
 748     assert(((ABS(iv_adjustment_in_bytes) % elt_size) == 0),
 749            err_msg_res("(%d) should be divisible by (%d)", iv_adjustment_in_bytes, elt_size));
 750     iv_adjustment = iv_adjustment_in_bytes/elt_size;
 751   } else {
 752     // This memory op is not dependent on iv (scale == 0)
 753     iv_adjustment = 0;
 754   }
 755 
 756 #ifndef PRODUCT
 757   if (TraceSuperWord)
 758     tty->print_cr("\noffset = %d iv_adjust = %d elt_size = %d scale = %d iv_stride = %d vect_size %d",
 759                   offset, iv_adjustment, elt_size, scale, iv_stride(), vw);
 760 #endif
 761   return iv_adjustment;
 762 }
 763 
 764 //---------------------------dependence_graph---------------------------
 765 // Construct dependency graph.
 766 // Add dependence edges to load/store nodes for memory dependence
 767 //    A.out()->DependNode.in(1) and DependNode.out()->B.prec(x)
 768 void SuperWord::dependence_graph() {
 769   // First, assign a dependence node to each memory node
 770   for (int i = 0; i < _block.length(); i++ ) {
 771     Node *n = _block.at(i);
 772     if (n->is_Mem() || n->is_Phi() && n->bottom_type() == Type::MEMORY) {
 773       _dg.make_node(n);
 774     }
 775   }
 776 
 777   // For each memory slice, create the dependences
 778   for (int i = 0; i < _mem_slice_head.length(); i++) {
 779     Node* n      = _mem_slice_head.at(i);
 780     Node* n_tail = _mem_slice_tail.at(i);
 781 
 782     // Get slice in predecessor order (last is first)
 783     mem_slice_preds(n_tail, n, _nlist);
 784 
 785 #ifndef PRODUCT
 786     if(TraceSuperWord && Verbose) {
 787       tty->print_cr("SuperWord::dependence_graph: built a new mem slice");
 788       for (int j = _nlist.length() - 1; j >= 0 ; j--) {
 789         _nlist.at(j)->dump();
 790       }
 791     }
 792 #endif
 793     // Make the slice dependent on the root
 794     DepMem* slice = _dg.dep(n);
 795     _dg.make_edge(_dg.root(), slice);
 796 
 797     // Create a sink for the slice
 798     DepMem* slice_sink = _dg.make_node(NULL);
 799     _dg.make_edge(slice_sink, _dg.tail());
 800 
 801     // Now visit each pair of memory ops, creating the edges
 802     for (int j = _nlist.length() - 1; j >= 0 ; j--) {
 803       Node* s1 = _nlist.at(j);
 804 
 805       // If no dependency yet, use slice
 806       if (_dg.dep(s1)->in_cnt() == 0) {
 807         _dg.make_edge(slice, s1);
 808       }
 809       SWPointer p1(s1->as_Mem(), this, NULL, false);
 810       bool sink_dependent = true;
 811       for (int k = j - 1; k >= 0; k--) {
 812         Node* s2 = _nlist.at(k);
 813         if (s1->is_Load() && s2->is_Load())
 814           continue;
 815         SWPointer p2(s2->as_Mem(), this, NULL, false);
 816 
 817         int cmp = p1.cmp(p2);
 818         if (SuperWordRTDepCheck &&
 819             p1.base() != p2.base() && p1.valid() && p2.valid()) {
 820           // Create a runtime check to disambiguate
 821           OrderedPair pp(p1.base(), p2.base());
 822           _disjoint_ptrs.append_if_missing(pp);
 823         } else if (!SWPointer::not_equal(cmp)) {
 824           // Possibly same address
 825           _dg.make_edge(s1, s2);
 826           sink_dependent = false;
 827         }
 828       }
 829       if (sink_dependent) {
 830         _dg.make_edge(s1, slice_sink);
 831       }
 832     }
 833 #ifndef PRODUCT
 834     if (TraceSuperWord) {
 835       tty->print_cr("\nDependence graph for slice: %d", n->_idx);
 836       for (int q = 0; q < _nlist.length(); q++) {
 837         _dg.print(_nlist.at(q));
 838       }
 839       tty->cr();
 840     }
 841 #endif
 842     _nlist.clear();
 843   }
 844 
 845 #ifndef PRODUCT
 846   if (TraceSuperWord) {
 847     tty->print_cr("\ndisjoint_ptrs: %s", _disjoint_ptrs.length() > 0 ? "" : "NONE");
 848     for (int r = 0; r < _disjoint_ptrs.length(); r++) {
 849       _disjoint_ptrs.at(r).print();
 850       tty->cr();
 851     }
 852     tty->cr();
 853   }
 854 #endif
 855 }
 856 
 857 //---------------------------mem_slice_preds---------------------------
 858 // Return a memory slice (node list) in predecessor order starting at "start"
 859 void SuperWord::mem_slice_preds(Node* start, Node* stop, GrowableArray<Node*> &preds) {
 860   assert(preds.length() == 0, "start empty");
 861   Node* n = start;
 862   Node* prev = NULL;
 863   while (true) {
 864     assert(in_bb(n), "must be in block");
 865     for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
 866       Node* out = n->fast_out(i);
 867       if (out->is_Load()) {
 868         if (in_bb(out)) {
 869           preds.push(out);
 870         }
 871       } else {
 872         // FIXME
 873         if (out->is_MergeMem() && !in_bb(out)) {
 874           // Either unrolling is causing a memory edge not to disappear,
 875           // or need to run igvn.optimize() again before SLP
 876         } else if (out->is_Phi() && out->bottom_type() == Type::MEMORY && !in_bb(out)) {
 877           // Ditto.  Not sure what else to check further.
 878         } else if (out->Opcode() == Op_StoreCM && out->in(MemNode::OopStore) == n) {
 879           // StoreCM has an input edge used as a precedence edge.
 880           // Maybe an issue when oop stores are vectorized.
 881         } else {
 882           assert(out == prev || prev == NULL, "no branches off of store slice");
 883         }
 884       }
 885     }
 886     if (n == stop) break;
 887     preds.push(n);
 888     prev = n;
 889     assert(n->is_Mem(), err_msg_res("unexpected node %s", n->Name()));
 890     n = n->in(MemNode::Memory);
 891   }
 892 }
 893 
 894 //------------------------------stmts_can_pack---------------------------
 895 // Can s1 and s2 be in a pack with s1 immediately preceding s2 and
 896 // s1 aligned at "align"
 897 bool SuperWord::stmts_can_pack(Node* s1, Node* s2, int align) {
 898 
 899   // Do not use superword for non-primitives
 900   BasicType bt1 = velt_basic_type(s1);
 901   BasicType bt2 = velt_basic_type(s2);
 902   if(!is_java_primitive(bt1) || !is_java_primitive(bt2))
 903     return false;
 904   if (Matcher::max_vector_size(bt1) < 2) {
 905     return false; // No vectors for this type
 906   }
 907 
 908   if (isomorphic(s1, s2)) {
 909     if (independent(s1, s2) || reduction(s1, s2)) {
 910       if (!exists_at(s1, 0) && !exists_at(s2, 1)) {
 911         if (!s1->is_Mem() || are_adjacent_refs(s1, s2)) {
 912           int s1_align = alignment(s1);
 913           int s2_align = alignment(s2);
 914           if (s1_align == top_align || s1_align == align) {
 915             if (s2_align == top_align || s2_align == align + data_size(s1)) {
 916               return true;
 917             }
 918           }
 919         }
 920       }
 921     }
 922   }
 923   return false;
 924 }
 925 
 926 //------------------------------exists_at---------------------------
 927 // Does s exist in a pack at position pos?
 928 bool SuperWord::exists_at(Node* s, uint pos) {
 929   for (int i = 0; i < _packset.length(); i++) {
 930     Node_List* p = _packset.at(i);
 931     if (p->at(pos) == s) {
 932       return true;
 933     }
 934   }
 935   return false;
 936 }
 937 
 938 //------------------------------are_adjacent_refs---------------------------
 939 // Is s1 immediately before s2 in memory?
 940 bool SuperWord::are_adjacent_refs(Node* s1, Node* s2) {
 941   if (!s1->is_Mem() || !s2->is_Mem()) return false;
 942   if (!in_bb(s1)    || !in_bb(s2))    return false;
 943 
 944   // Do not use superword for non-primitives
 945   if (!is_java_primitive(s1->as_Mem()->memory_type()) ||
 946       !is_java_primitive(s2->as_Mem()->memory_type())) {
 947     return false;
 948   }
 949 
 950   // FIXME - co_locate_pack fails on Stores in different mem-slices, so
 951   // only pack memops that are in the same alias set until that's fixed.
 952   if (_phase->C->get_alias_index(s1->as_Mem()->adr_type()) !=
 953       _phase->C->get_alias_index(s2->as_Mem()->adr_type()))
 954     return false;
 955   SWPointer p1(s1->as_Mem(), this, NULL, false);
 956   SWPointer p2(s2->as_Mem(), this, NULL, false);
 957   if (p1.base() != p2.base() || !p1.comparable(p2)) return false;
 958   int diff = p2.offset_in_bytes() - p1.offset_in_bytes();
 959   return diff == data_size(s1);
 960 }
 961 
 962 //------------------------------isomorphic---------------------------
 963 // Are s1 and s2 similar?
 964 bool SuperWord::isomorphic(Node* s1, Node* s2) {
 965   if (s1->Opcode() != s2->Opcode()) return false;
 966   if (s1->req() != s2->req()) return false;
 967   if (s1->in(0) != s2->in(0)) return false;
 968   if (!same_velt_type(s1, s2)) return false;
 969   return true;
 970 }
 971 
 972 //------------------------------independent---------------------------
 973 // Is there no data path from s1 to s2 or s2 to s1?
 974 bool SuperWord::independent(Node* s1, Node* s2) {
 975   //  assert(s1->Opcode() == s2->Opcode(), "check isomorphic first");
 976   int d1 = depth(s1);
 977   int d2 = depth(s2);
 978   if (d1 == d2) return s1 != s2;
 979   Node* deep    = d1 > d2 ? s1 : s2;
 980   Node* shallow = d1 > d2 ? s2 : s1;
 981 
 982   visited_clear();
 983 
 984   return independent_path(shallow, deep);
 985 }
 986 
 987 //------------------------------reduction---------------------------
 988 // Is there a data path between s1 and s2 and the nodes reductions?
 989 bool SuperWord::reduction(Node* s1, Node* s2) {
 990   bool retValue = false;
 991   int d1 = depth(s1);
 992   int d2 = depth(s2);
 993   if (d1 + 1 == d2) {
 994     if (s1->is_reduction() && s2->is_reduction()) {
 995       // This is an ordered set, so s1 should define s2
 996       for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
 997         Node* t1 = s1->fast_out(i);
 998         if (t1 == s2) {
 999           // both nodes are reductions and connected
1000           retValue = true;
1001         }
1002       }
1003     }
1004   }
1005 
1006   return retValue;
1007 }
1008 
1009 //------------------------------independent_path------------------------------
1010 // Helper for independent
1011 bool SuperWord::independent_path(Node* shallow, Node* deep, uint dp) {
1012   if (dp >= 1000) return false; // stop deep recursion
1013   visited_set(deep);
1014   int shal_depth = depth(shallow);
1015   assert(shal_depth <= depth(deep), "must be");
1016   for (DepPreds preds(deep, _dg); !preds.done(); preds.next()) {
1017     Node* pred = preds.current();
1018     if (in_bb(pred) && !visited_test(pred)) {
1019       if (shallow == pred) {
1020         return false;
1021       }
1022       if (shal_depth < depth(pred) && !independent_path(shallow, pred, dp+1)) {
1023         return false;
1024       }
1025     }
1026   }
1027   return true;
1028 }
1029 
1030 //------------------------------set_alignment---------------------------
1031 void SuperWord::set_alignment(Node* s1, Node* s2, int align) {
1032   set_alignment(s1, align);
1033   if (align == top_align || align == bottom_align) {
1034     set_alignment(s2, align);
1035   } else {
1036     set_alignment(s2, align + data_size(s1));
1037   }
1038 }
1039 
1040 //------------------------------data_size---------------------------
1041 int SuperWord::data_size(Node* s) {
1042   int bsize = type2aelembytes(velt_basic_type(s));
1043   assert(bsize != 0, "valid size");
1044   return bsize;
1045 }
1046 
1047 //------------------------------extend_packlist---------------------------
1048 // Extend packset by following use->def and def->use links from pack members.
1049 void SuperWord::extend_packlist() {
1050   bool changed;
1051   do {
1052     packset_sort(_packset.length());
1053     changed = false;
1054     for (int i = 0; i < _packset.length(); i++) {
1055       Node_List* p = _packset.at(i);
1056       changed |= follow_use_defs(p);
1057       changed |= follow_def_uses(p);
1058     }
1059   } while (changed);
1060 
1061   if (_race_possible) {
1062     for (int i = 0; i < _packset.length(); i++) {
1063       Node_List* p = _packset.at(i);
1064       order_def_uses(p);
1065     }
1066   }
1067 
1068 #ifndef PRODUCT
1069   if (TraceSuperWord) {
1070     tty->print_cr("\nAfter extend_packlist");
1071     print_packset();
1072   }
1073 #endif
1074 }
1075 
1076 //------------------------------follow_use_defs---------------------------
1077 // Extend the packset by visiting operand definitions of nodes in pack p
1078 bool SuperWord::follow_use_defs(Node_List* p) {
1079   assert(p->size() == 2, "just checking");
1080   Node* s1 = p->at(0);
1081   Node* s2 = p->at(1);
1082   assert(s1->req() == s2->req(), "just checking");
1083   assert(alignment(s1) + data_size(s1) == alignment(s2), "just checking");
1084 
1085   if (s1->is_Load()) return false;
1086 
1087   int align = alignment(s1);
1088   bool changed = false;
1089   int start = s1->is_Store() ? MemNode::ValueIn   : 1;
1090   int end   = s1->is_Store() ? MemNode::ValueIn+1 : s1->req();
1091   for (int j = start; j < end; j++) {
1092     Node* t1 = s1->in(j);
1093     Node* t2 = s2->in(j);
1094     if (!in_bb(t1) || !in_bb(t2))
1095       continue;
1096     if (stmts_can_pack(t1, t2, align)) {
1097       if (est_savings(t1, t2) >= 0) {
1098         Node_List* pair = new Node_List();
1099         pair->push(t1);
1100         pair->push(t2);
1101         _packset.append(pair);
1102         set_alignment(t1, t2, align);
1103         changed = true;
1104       }
1105     }
1106   }
1107   return changed;
1108 }
1109 
1110 //------------------------------follow_def_uses---------------------------
1111 // Extend the packset by visiting uses of nodes in pack p
1112 bool SuperWord::follow_def_uses(Node_List* p) {
1113   bool changed = false;
1114   Node* s1 = p->at(0);
1115   Node* s2 = p->at(1);
1116   assert(p->size() == 2, "just checking");
1117   assert(s1->req() == s2->req(), "just checking");
1118   assert(alignment(s1) + data_size(s1) == alignment(s2), "just checking");
1119 
1120   if (s1->is_Store()) return false;
1121 
1122   int align = alignment(s1);
1123   int savings = -1;
1124   int num_s1_uses = 0;
1125   Node* u1 = NULL;
1126   Node* u2 = NULL;
1127   for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
1128     Node* t1 = s1->fast_out(i);
1129     num_s1_uses++;
1130     if (!in_bb(t1)) continue;
1131     for (DUIterator_Fast jmax, j = s2->fast_outs(jmax); j < jmax; j++) {
1132       Node* t2 = s2->fast_out(j);
1133       if (!in_bb(t2)) continue;
1134       if (!opnd_positions_match(s1, t1, s2, t2))
1135         continue;
1136       if (stmts_can_pack(t1, t2, align)) {
1137         int my_savings = est_savings(t1, t2);
1138         if (my_savings > savings) {
1139           savings = my_savings;
1140           u1 = t1;
1141           u2 = t2;
1142         }
1143       }
1144     }
1145   }
1146   if (num_s1_uses > 1) {
1147     _race_possible = true;
1148   }
1149   if (savings >= 0) {
1150     Node_List* pair = new Node_List();
1151     pair->push(u1);
1152     pair->push(u2);
1153     _packset.append(pair);
1154     set_alignment(u1, u2, align);
1155     changed = true;
1156   }
1157   return changed;
1158 }
1159 
1160 //------------------------------order_def_uses---------------------------
1161 // For extended packsets, ordinally arrange uses packset by major component
1162 void SuperWord::order_def_uses(Node_List* p) {
1163   Node* s1 = p->at(0);
1164 
1165   if (s1->is_Store()) return;
1166 
1167   // reductions are always managed beforehand
1168   if (s1->is_reduction()) return;
1169 
1170   for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
1171     Node* t1 = s1->fast_out(i);
1172 
1173     // Only allow operand swap on commuting operations
1174     if (!t1->is_Add() && !t1->is_Mul()) {
1175       break;
1176     }
1177 
1178     // Now find t1's packset
1179     Node_List* p2 = NULL;
1180     for (int j = 0; j < _packset.length(); j++) {
1181       p2 = _packset.at(j);
1182       Node* first = p2->at(0);
1183       if (t1 == first) {
1184         break;
1185       }
1186       p2 = NULL;
1187     }
1188     // Arrange all sub components by the major component
1189     if (p2 != NULL) {
1190       for (uint j = 1; j < p->size(); j++) {
1191         Node* d1 = p->at(j);
1192         Node* u1 = p2->at(j);
1193         opnd_positions_match(s1, t1, d1, u1);
1194       }
1195     }
1196   }
1197 }
1198 
1199 //---------------------------opnd_positions_match-------------------------
1200 // Is the use of d1 in u1 at the same operand position as d2 in u2?
1201 bool SuperWord::opnd_positions_match(Node* d1, Node* u1, Node* d2, Node* u2) {
1202   // check reductions to see if they are marshalled to represent the reduction
1203   // operator in a specified opnd
1204   if (u1->is_reduction() && u2->is_reduction()) {
1205     // ensure reductions have phis and reduction definitions feeding the 1st operand
1206     Node* first = u1->in(2);
1207     if (first->is_Phi() || first->is_reduction()) {
1208       u1->swap_edges(1, 2);
1209     }
1210     // ensure reductions have phis and reduction definitions feeding the 1st operand
1211     first = u2->in(2);
1212     if (first->is_Phi() || first->is_reduction()) {
1213       u2->swap_edges(1, 2);
1214     }
1215     return true;
1216   }
1217 
1218   uint ct = u1->req();
1219   if (ct != u2->req()) return false;
1220   uint i1 = 0;
1221   uint i2 = 0;
1222   do {
1223     for (i1++; i1 < ct; i1++) if (u1->in(i1) == d1) break;
1224     for (i2++; i2 < ct; i2++) if (u2->in(i2) == d2) break;
1225     if (i1 != i2) {
1226       if ((i1 == (3-i2)) && (u2->is_Add() || u2->is_Mul())) {
1227         // Further analysis relies on operands position matching.
1228         u2->swap_edges(i1, i2);
1229       } else {
1230         return false;
1231       }
1232     }
1233   } while (i1 < ct);
1234   return true;
1235 }
1236 
1237 //------------------------------est_savings---------------------------
1238 // Estimate the savings from executing s1 and s2 as a pack
1239 int SuperWord::est_savings(Node* s1, Node* s2) {
1240   int save_in = 2 - 1; // 2 operations per instruction in packed form
1241 
1242   // inputs
1243   for (uint i = 1; i < s1->req(); i++) {
1244     Node* x1 = s1->in(i);
1245     Node* x2 = s2->in(i);
1246     if (x1 != x2) {
1247       if (are_adjacent_refs(x1, x2)) {
1248         save_in += adjacent_profit(x1, x2);
1249       } else if (!in_packset(x1, x2)) {
1250         save_in -= pack_cost(2);
1251       } else {
1252         save_in += unpack_cost(2);
1253       }
1254     }
1255   }
1256 
1257   // uses of result
1258   uint ct = 0;
1259   int save_use = 0;
1260   for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
1261     Node* s1_use = s1->fast_out(i);
1262     for (int j = 0; j < _packset.length(); j++) {
1263       Node_List* p = _packset.at(j);
1264       if (p->at(0) == s1_use) {
1265         for (DUIterator_Fast kmax, k = s2->fast_outs(kmax); k < kmax; k++) {
1266           Node* s2_use = s2->fast_out(k);
1267           if (p->at(p->size()-1) == s2_use) {
1268             ct++;
1269             if (are_adjacent_refs(s1_use, s2_use)) {
1270               save_use += adjacent_profit(s1_use, s2_use);
1271             }
1272           }
1273         }
1274       }
1275     }
1276   }
1277 
1278   if (ct < s1->outcnt()) save_use += unpack_cost(1);
1279   if (ct < s2->outcnt()) save_use += unpack_cost(1);
1280 
1281   return MAX2(save_in, save_use);
1282 }
1283 
1284 //------------------------------costs---------------------------
1285 int SuperWord::adjacent_profit(Node* s1, Node* s2) { return 2; }
1286 int SuperWord::pack_cost(int ct)   { return ct; }
1287 int SuperWord::unpack_cost(int ct) { return ct; }
1288 
1289 //------------------------------combine_packs---------------------------
1290 // Combine packs A and B with A.last == B.first into A.first..,A.last,B.second,..B.last
1291 void SuperWord::combine_packs() {
1292   bool changed = true;
1293   // Combine packs regardless max vector size.
1294   while (changed) {
1295     changed = false;
1296     for (int i = 0; i < _packset.length(); i++) {
1297       Node_List* p1 = _packset.at(i);
1298       if (p1 == NULL) continue;
1299       // Because of sorting we can start at i + 1
1300       for (int j = i + 1; j < _packset.length(); j++) {
1301         Node_List* p2 = _packset.at(j);
1302         if (p2 == NULL) continue;
1303         if (i == j) continue;
1304         if (p1->at(p1->size()-1) == p2->at(0)) {
1305           for (uint k = 1; k < p2->size(); k++) {
1306             p1->push(p2->at(k));
1307           }
1308           _packset.at_put(j, NULL);
1309           changed = true;
1310         }
1311       }
1312     }
1313   }
1314 
1315   // Split packs which have size greater then max vector size.
1316   for (int i = 0; i < _packset.length(); i++) {
1317     Node_List* p1 = _packset.at(i);
1318     if (p1 != NULL) {
1319       BasicType bt = velt_basic_type(p1->at(0));
1320       uint max_vlen = Matcher::max_vector_size(bt); // Max elements in vector
1321       assert(is_power_of_2(max_vlen), "sanity");
1322       uint psize = p1->size();
1323       if (!is_power_of_2(psize)) {
1324         // Skip pack which can't be vector.
1325         // case1: for(...) { a[i] = i; }    elements values are different (i+x)
1326         // case2: for(...) { a[i] = b[i+1]; }  can't align both, load and store
1327         _packset.at_put(i, NULL);
1328         continue;
1329       }
1330       if (psize > max_vlen) {
1331         Node_List* pack = new Node_List();
1332         for (uint j = 0; j < psize; j++) {
1333           pack->push(p1->at(j));
1334           if (pack->size() >= max_vlen) {
1335             assert(is_power_of_2(pack->size()), "sanity");
1336             _packset.append(pack);
1337             pack = new Node_List();
1338           }
1339         }
1340         _packset.at_put(i, NULL);
1341       }
1342     }
1343   }
1344 
1345   // Compress list.
1346   for (int i = _packset.length() - 1; i >= 0; i--) {
1347     Node_List* p1 = _packset.at(i);
1348     if (p1 == NULL) {
1349       _packset.remove_at(i);
1350     }
1351   }
1352 
1353 #ifndef PRODUCT
1354   if (TraceSuperWord) {
1355     tty->print_cr("\nAfter combine_packs");
1356     print_packset();
1357   }
1358 #endif
1359 }
1360 
1361 //-----------------------------construct_my_pack_map--------------------------
1362 // Construct the map from nodes to packs.  Only valid after the
1363 // point where a node is only in one pack (after combine_packs).
1364 void SuperWord::construct_my_pack_map() {
1365   Node_List* rslt = NULL;
1366   for (int i = 0; i < _packset.length(); i++) {
1367     Node_List* p = _packset.at(i);
1368     for (uint j = 0; j < p->size(); j++) {
1369       Node* s = p->at(j);
1370       assert(my_pack(s) == NULL, "only in one pack");
1371       set_my_pack(s, p);
1372     }
1373   }
1374 }
1375 
1376 //------------------------------filter_packs---------------------------
1377 // Remove packs that are not implemented or not profitable.
1378 void SuperWord::filter_packs() {
1379   // Remove packs that are not implemented
1380   for (int i = _packset.length() - 1; i >= 0; i--) {
1381     Node_List* pk = _packset.at(i);
1382     bool impl = implemented(pk);
1383     if (!impl) {
1384 #ifndef PRODUCT
1385       if (TraceSuperWord && Verbose) {
1386         tty->print_cr("Unimplemented");
1387         pk->at(0)->dump();
1388       }
1389 #endif
1390       remove_pack_at(i);
1391     }
1392     Node *n = pk->at(0);
1393     if (n->is_reduction()) {
1394       _num_reductions++;
1395     } else {
1396       _num_work_vecs++;
1397     }
1398   }
1399 
1400   // Remove packs that are not profitable
1401   bool changed;
1402   do {
1403     changed = false;
1404     for (int i = _packset.length() - 1; i >= 0; i--) {
1405       Node_List* pk = _packset.at(i);
1406       bool prof = profitable(pk);
1407       if (!prof) {
1408 #ifndef PRODUCT
1409         if (TraceSuperWord && Verbose) {
1410           tty->print_cr("Unprofitable");
1411           pk->at(0)->dump();
1412         }
1413 #endif
1414         remove_pack_at(i);
1415         changed = true;
1416       }
1417     }
1418   } while (changed);
1419 
1420 #ifndef PRODUCT
1421   if (TraceSuperWord) {
1422     tty->print_cr("\nAfter filter_packs");
1423     print_packset();
1424     tty->cr();
1425   }
1426 #endif
1427 }
1428 
1429 //------------------------------implemented---------------------------
1430 // Can code be generated for pack p?
1431 bool SuperWord::implemented(Node_List* p) {
1432   bool retValue = false;
1433   Node* p0 = p->at(0);
1434   if (p0 != NULL) {
1435     int opc = p0->Opcode();
1436     uint size = p->size();
1437     if (p0->is_reduction()) {
1438       const Type *arith_type = p0->bottom_type();
1439       // Length 2 reductions of INT/LONG do not offer performance benefits
1440       if (((arith_type->basic_type() == T_INT) || (arith_type->basic_type() == T_LONG)) && (size == 2)) {
1441         retValue = false;
1442       } else {
1443         retValue = ReductionNode::implemented(opc, size, arith_type->basic_type());
1444       }
1445     } else {
1446       retValue = VectorNode::implemented(opc, size, velt_basic_type(p0));
1447     }
1448   }
1449   return retValue;
1450 }
1451 
1452 //------------------------------same_inputs--------------------------
1453 // For pack p, are all idx operands the same?
1454 static bool same_inputs(Node_List* p, int idx) {
1455   Node* p0 = p->at(0);
1456   uint vlen = p->size();
1457   Node* p0_def = p0->in(idx);
1458   for (uint i = 1; i < vlen; i++) {
1459     Node* pi = p->at(i);
1460     Node* pi_def = pi->in(idx);
1461     if (p0_def != pi_def)
1462       return false;
1463   }
1464   return true;
1465 }
1466 
1467 //------------------------------profitable---------------------------
1468 // For pack p, are all operands and all uses (with in the block) vector?
1469 bool SuperWord::profitable(Node_List* p) {
1470   Node* p0 = p->at(0);
1471   uint start, end;
1472   VectorNode::vector_operands(p0, &start, &end);
1473 
1474   // Return false if some inputs are not vectors or vectors with different
1475   // size or alignment.
1476   // Also, for now, return false if not scalar promotion case when inputs are
1477   // the same. Later, implement PackNode and allow differing, non-vector inputs
1478   // (maybe just the ones from outside the block.)
1479   for (uint i = start; i < end; i++) {
1480     if (!is_vector_use(p0, i))
1481       return false;
1482   }
1483   // Check if reductions are connected
1484   if (p0->is_reduction()) {
1485     Node* second_in = p0->in(2);
1486     Node_List* second_pk = my_pack(second_in);
1487     if ((second_pk == NULL) || (_num_work_vecs == _num_reductions)) {
1488       // Remove reduction flag if no parent pack or if not enough work
1489       // to cover reduction expansion overhead
1490       p0->remove_flag(Node::Flag_is_reduction);
1491       return false;
1492     } else if (second_pk->size() != p->size()) {
1493       return false;
1494     }
1495   }
1496   if (VectorNode::is_shift(p0)) {
1497     // For now, return false if shift count is vector or not scalar promotion
1498     // case (different shift counts) because it is not supported yet.
1499     Node* cnt = p0->in(2);
1500     Node_List* cnt_pk = my_pack(cnt);
1501     if (cnt_pk != NULL)
1502       return false;
1503     if (!same_inputs(p, 2))
1504       return false;
1505   }
1506   if (!p0->is_Store()) {
1507     // For now, return false if not all uses are vector.
1508     // Later, implement ExtractNode and allow non-vector uses (maybe
1509     // just the ones outside the block.)
1510     for (uint i = 0; i < p->size(); i++) {
1511       Node* def = p->at(i);
1512       for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
1513         Node* use = def->fast_out(j);
1514         for (uint k = 0; k < use->req(); k++) {
1515           Node* n = use->in(k);
1516           if (def == n) {
1517             // reductions can be loop carried dependences
1518             if (def->is_reduction() && use->is_Phi())
1519               continue;
1520             if (!is_vector_use(use, k)) {
1521               return false;
1522             }
1523           }
1524         }
1525       }
1526     }
1527   }
1528   return true;
1529 }
1530 
1531 //------------------------------schedule---------------------------
1532 // Adjust the memory graph for the packed operations
1533 void SuperWord::schedule() {
1534 
1535   // Co-locate in the memory graph the members of each memory pack
1536   for (int i = 0; i < _packset.length(); i++) {
1537     co_locate_pack(_packset.at(i));
1538   }
1539 }
1540 
1541 //-------------------------------remove_and_insert-------------------
1542 // Remove "current" from its current position in the memory graph and insert
1543 // it after the appropriate insertion point (lip or uip).
1544 void SuperWord::remove_and_insert(MemNode *current, MemNode *prev, MemNode *lip,
1545                                   Node *uip, Unique_Node_List &sched_before) {
1546   Node* my_mem = current->in(MemNode::Memory);
1547   bool sched_up = sched_before.member(current);
1548 
1549   // remove current_store from its current position in the memmory graph
1550   for (DUIterator i = current->outs(); current->has_out(i); i++) {
1551     Node* use = current->out(i);
1552     if (use->is_Mem()) {
1553       assert(use->in(MemNode::Memory) == current, "must be");
1554       if (use == prev) { // connect prev to my_mem
1555           _igvn.replace_input_of(use, MemNode::Memory, my_mem);
1556           --i; //deleted this edge; rescan position
1557       } else if (sched_before.member(use)) {
1558         if (!sched_up) { // Will be moved together with current
1559           _igvn.replace_input_of(use, MemNode::Memory, uip);
1560           --i; //deleted this edge; rescan position
1561         }
1562       } else {
1563         if (sched_up) { // Will be moved together with current
1564           _igvn.replace_input_of(use, MemNode::Memory, lip);
1565           --i; //deleted this edge; rescan position
1566         }
1567       }
1568     }
1569   }
1570 
1571   Node *insert_pt =  sched_up ?  uip : lip;
1572 
1573   // all uses of insert_pt's memory state should use current's instead
1574   for (DUIterator i = insert_pt->outs(); insert_pt->has_out(i); i++) {
1575     Node* use = insert_pt->out(i);
1576     if (use->is_Mem()) {
1577       assert(use->in(MemNode::Memory) == insert_pt, "must be");
1578       _igvn.replace_input_of(use, MemNode::Memory, current);
1579       --i; //deleted this edge; rescan position
1580     } else if (!sched_up && use->is_Phi() && use->bottom_type() == Type::MEMORY) {
1581       uint pos; //lip (lower insert point) must be the last one in the memory slice
1582       for (pos=1; pos < use->req(); pos++) {
1583         if (use->in(pos) == insert_pt) break;
1584       }
1585       _igvn.replace_input_of(use, pos, current);
1586       --i;
1587     }
1588   }
1589 
1590   //connect current to insert_pt
1591   _igvn.replace_input_of(current, MemNode::Memory, insert_pt);
1592 }
1593 
1594 //------------------------------co_locate_pack----------------------------------
1595 // To schedule a store pack, we need to move any sandwiched memory ops either before
1596 // or after the pack, based upon dependence information:
1597 // (1) If any store in the pack depends on the sandwiched memory op, the
1598 //     sandwiched memory op must be scheduled BEFORE the pack;
1599 // (2) If a sandwiched memory op depends on any store in the pack, the
1600 //     sandwiched memory op must be scheduled AFTER the pack;
1601 // (3) If a sandwiched memory op (say, memA) depends on another sandwiched
1602 //     memory op (say memB), memB must be scheduled before memA. So, if memA is
1603 //     scheduled before the pack, memB must also be scheduled before the pack;
1604 // (4) If there is no dependence restriction for a sandwiched memory op, we simply
1605 //     schedule this store AFTER the pack
1606 // (5) We know there is no dependence cycle, so there in no other case;
1607 // (6) Finally, all memory ops in another single pack should be moved in the same direction.
1608 //
1609 // To schedule a load pack, we use the memory state of either the first or the last load in
1610 // the pack, based on the dependence constraint.
1611 void SuperWord::co_locate_pack(Node_List* pk) {
1612   if (pk->at(0)->is_Store()) {
1613     MemNode* first     = executed_first(pk)->as_Mem();
1614     MemNode* last      = executed_last(pk)->as_Mem();
1615     Unique_Node_List schedule_before_pack;
1616     Unique_Node_List memops;
1617 
1618     MemNode* current   = last->in(MemNode::Memory)->as_Mem();
1619     MemNode* previous  = last;
1620     while (true) {
1621       assert(in_bb(current), "stay in block");
1622       memops.push(previous);
1623       for (DUIterator i = current->outs(); current->has_out(i); i++) {
1624         Node* use = current->out(i);
1625         if (use->is_Mem() && use != previous)
1626           memops.push(use);
1627       }
1628       if (current == first) break;
1629       previous = current;
1630       current  = current->in(MemNode::Memory)->as_Mem();
1631     }
1632 
1633     // determine which memory operations should be scheduled before the pack
1634     for (uint i = 1; i < memops.size(); i++) {
1635       Node *s1 = memops.at(i);
1636       if (!in_pack(s1, pk) && !schedule_before_pack.member(s1)) {
1637         for (uint j = 0; j< i; j++) {
1638           Node *s2 = memops.at(j);
1639           if (!independent(s1, s2)) {
1640             if (in_pack(s2, pk) || schedule_before_pack.member(s2)) {
1641               schedule_before_pack.push(s1); // s1 must be scheduled before
1642               Node_List* mem_pk = my_pack(s1);
1643               if (mem_pk != NULL) {
1644                 for (uint ii = 0; ii < mem_pk->size(); ii++) {
1645                   Node* s = mem_pk->at(ii);  // follow partner
1646                   if (memops.member(s) && !schedule_before_pack.member(s))
1647                     schedule_before_pack.push(s);
1648                 }
1649               }
1650               break;
1651             }
1652           }
1653         }
1654       }
1655     }
1656 
1657     Node*    upper_insert_pt = first->in(MemNode::Memory);
1658     // Following code moves loads connected to upper_insert_pt below aliased stores.
1659     // Collect such loads here and reconnect them back to upper_insert_pt later.
1660     memops.clear();
1661     for (DUIterator i = upper_insert_pt->outs(); upper_insert_pt->has_out(i); i++) {
1662       Node* use = upper_insert_pt->out(i);
1663       if (use->is_Mem() && !use->is_Store()) {
1664         memops.push(use);
1665       }
1666     }
1667 
1668     MemNode* lower_insert_pt = last;
1669     previous                 = last; //previous store in pk
1670     current                  = last->in(MemNode::Memory)->as_Mem();
1671 
1672     // start scheduling from "last" to "first"
1673     while (true) {
1674       assert(in_bb(current), "stay in block");
1675       assert(in_pack(previous, pk), "previous stays in pack");
1676       Node* my_mem = current->in(MemNode::Memory);
1677 
1678       if (in_pack(current, pk)) {
1679         // Forward users of my memory state (except "previous) to my input memory state
1680         for (DUIterator i = current->outs(); current->has_out(i); i++) {
1681           Node* use = current->out(i);
1682           if (use->is_Mem() && use != previous) {
1683             assert(use->in(MemNode::Memory) == current, "must be");
1684             if (schedule_before_pack.member(use)) {
1685               _igvn.replace_input_of(use, MemNode::Memory, upper_insert_pt);
1686             } else {
1687               _igvn.replace_input_of(use, MemNode::Memory, lower_insert_pt);
1688             }
1689             --i; // deleted this edge; rescan position
1690           }
1691         }
1692         previous = current;
1693       } else { // !in_pack(current, pk) ==> a sandwiched store
1694         remove_and_insert(current, previous, lower_insert_pt, upper_insert_pt, schedule_before_pack);
1695       }
1696 
1697       if (current == first) break;
1698       current = my_mem->as_Mem();
1699     } // end while
1700 
1701     // Reconnect loads back to upper_insert_pt.
1702     for (uint i = 0; i < memops.size(); i++) {
1703       Node *ld = memops.at(i);
1704       if (ld->in(MemNode::Memory) != upper_insert_pt) {
1705         _igvn.replace_input_of(ld, MemNode::Memory, upper_insert_pt);
1706       }
1707     }
1708   } else if (pk->at(0)->is_Load()) { //load
1709     // all loads in the pack should have the same memory state. By default,
1710     // we use the memory state of the last load. However, if any load could
1711     // not be moved down due to the dependence constraint, we use the memory
1712     // state of the first load.
1713     Node* last_mem  = executed_last(pk)->in(MemNode::Memory);
1714     Node* first_mem = executed_first(pk)->in(MemNode::Memory);
1715     bool schedule_last = true;
1716     for (uint i = 0; i < pk->size(); i++) {
1717       Node* ld = pk->at(i);
1718       for (Node* current = last_mem; current != ld->in(MemNode::Memory);
1719            current=current->in(MemNode::Memory)) {
1720         assert(current != first_mem, "corrupted memory graph");
1721         if(current->is_Mem() && !independent(current, ld)){
1722           schedule_last = false; // a later store depends on this load
1723           break;
1724         }
1725       }
1726     }
1727 
1728     Node* mem_input = schedule_last ? last_mem : first_mem;
1729     _igvn.hash_delete(mem_input);
1730     // Give each load the same memory state
1731     for (uint i = 0; i < pk->size(); i++) {
1732       LoadNode* ld = pk->at(i)->as_Load();
1733       _igvn.replace_input_of(ld, MemNode::Memory, mem_input);
1734     }
1735   }
1736 }
1737 
1738 //------------------------------output---------------------------
1739 // Convert packs into vector node operations
1740 void SuperWord::output() {
1741   if (_packset.length() == 0) return;
1742 
1743 #ifndef PRODUCT
1744   if (TraceLoopOpts) {
1745     tty->print("SuperWord    ");
1746     lpt()->dump_head();
1747   }
1748 #endif
1749 
1750   // MUST ENSURE main loop's initial value is properly aligned:
1751   //  (iv_initial_value + min_iv_offset) % vector_width_in_bytes() == 0
1752 
1753   align_initial_loop_index(align_to_ref());
1754 
1755   // Insert extract (unpack) operations for scalar uses
1756   for (int i = 0; i < _packset.length(); i++) {
1757     insert_extracts(_packset.at(i));
1758   }
1759 
1760   Compile* C = _phase->C;
1761   uint max_vlen_in_bytes = 0;
1762   for (int i = 0; i < _block.length(); i++) {
1763     Node* n = _block.at(i);
1764     Node_List* p = my_pack(n);
1765     if (p && n == executed_last(p)) {
1766       uint vlen = p->size();
1767       uint vlen_in_bytes = 0;
1768       Node* vn = NULL;
1769       Node* low_adr = p->at(0);
1770       Node* first   = executed_first(p);
1771       int   opc = n->Opcode();
1772       if (n->is_Load()) {
1773         Node* ctl = n->in(MemNode::Control);
1774         Node* mem = first->in(MemNode::Memory);
1775         SWPointer p1(n->as_Mem(), this, NULL, false);
1776         // Identify the memory dependency for the new loadVector node by
1777         // walking up through memory chain.
1778         // This is done to give flexibility to the new loadVector node so that
1779         // it can move above independent storeVector nodes.
1780         while (mem->is_StoreVector()) {
1781           SWPointer p2(mem->as_Mem(), this, NULL, false);
1782           int cmp = p1.cmp(p2);
1783           if (SWPointer::not_equal(cmp) || !SWPointer::comparable(cmp)) {
1784             mem = mem->in(MemNode::Memory);
1785           } else {
1786             break; // dependent memory
1787           }
1788         }
1789         Node* adr = low_adr->in(MemNode::Address);
1790         const TypePtr* atyp = n->adr_type();
1791         vn = LoadVectorNode::make(opc, ctl, mem, adr, atyp, vlen, velt_basic_type(n), control_dependency(p));
1792         vlen_in_bytes = vn->as_LoadVector()->memory_size();
1793       } else if (n->is_Store()) {
1794         // Promote value to be stored to vector
1795         Node* val = vector_opd(p, MemNode::ValueIn);
1796         Node* ctl = n->in(MemNode::Control);
1797         Node* mem = first->in(MemNode::Memory);
1798         Node* adr = low_adr->in(MemNode::Address);
1799         const TypePtr* atyp = n->adr_type();
1800         vn = StoreVectorNode::make(opc, ctl, mem, adr, atyp, val, vlen);
1801         vlen_in_bytes = vn->as_StoreVector()->memory_size();
1802       } else if (n->req() == 3) {
1803         // Promote operands to vector
1804         Node* in1 = NULL;
1805         bool node_isa_reduction = n->is_reduction();
1806         if (node_isa_reduction) {
1807           // the input to the first reduction operation is retained
1808           in1 = low_adr->in(1);
1809         } else {
1810           in1 = vector_opd(p, 1);
1811         }
1812         Node* in2 = vector_opd(p, 2);
1813         if (VectorNode::is_invariant_vector(in1) && (node_isa_reduction == false) && (n->is_Add() || n->is_Mul())) {
1814           // Move invariant vector input into second position to avoid register spilling.
1815           Node* tmp = in1;
1816           in1 = in2;
1817           in2 = tmp;
1818         }
1819         if (node_isa_reduction) {
1820           const Type *arith_type = n->bottom_type();
1821           vn = ReductionNode::make(opc, NULL, in1, in2, arith_type->basic_type());
1822           if (in2->is_Load()) {
1823             vlen_in_bytes = in2->as_LoadVector()->memory_size();
1824           } else {
1825             vlen_in_bytes = in2->as_Vector()->length_in_bytes();
1826           }
1827         } else {
1828           vn = VectorNode::make(opc, in1, in2, vlen, velt_basic_type(n));
1829           vlen_in_bytes = vn->as_Vector()->length_in_bytes();
1830         }
1831       } else {
1832         ShouldNotReachHere();
1833       }
1834       assert(vn != NULL, "sanity");
1835       _igvn.register_new_node_with_optimizer(vn);
1836       _phase->set_ctrl(vn, _phase->get_ctrl(p->at(0)));
1837       for (uint j = 0; j < p->size(); j++) {
1838         Node* pm = p->at(j);
1839         _igvn.replace_node(pm, vn);
1840       }
1841       _igvn._worklist.push(vn);
1842 
1843       if (vlen_in_bytes > max_vlen_in_bytes) {
1844         max_vlen_in_bytes = vlen_in_bytes;
1845       }
1846 #ifdef ASSERT
1847       if (TraceNewVectors) {
1848         tty->print("new Vector node: ");
1849         vn->dump();
1850       }
1851 #endif
1852     }
1853   }
1854   C->set_max_vector_size(max_vlen_in_bytes);
1855 }
1856 
1857 //------------------------------vector_opd---------------------------
1858 // Create a vector operand for the nodes in pack p for operand: in(opd_idx)
1859 Node* SuperWord::vector_opd(Node_List* p, int opd_idx) {
1860   Node* p0 = p->at(0);
1861   uint vlen = p->size();
1862   Node* opd = p0->in(opd_idx);
1863 
1864   if (same_inputs(p, opd_idx)) {
1865     if (opd->is_Vector() || opd->is_LoadVector()) {
1866       assert(((opd_idx != 2) || !VectorNode::is_shift(p0)), "shift's count can't be vector");
1867       return opd; // input is matching vector
1868     }
1869     if ((opd_idx == 2) && VectorNode::is_shift(p0)) {
1870       Compile* C = _phase->C;
1871       Node* cnt = opd;
1872       // Vector instructions do not mask shift count, do it here.
1873       juint mask = (p0->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
1874       const TypeInt* t = opd->find_int_type();
1875       if (t != NULL && t->is_con()) {
1876         juint shift = t->get_con();
1877         if (shift > mask) { // Unsigned cmp
1878           cnt = ConNode::make(TypeInt::make(shift & mask));
1879         }
1880       } else {
1881         if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
1882           cnt = ConNode::make(TypeInt::make(mask));
1883           _igvn.register_new_node_with_optimizer(cnt);
1884           cnt = new AndINode(opd, cnt);
1885           _igvn.register_new_node_with_optimizer(cnt);
1886           _phase->set_ctrl(cnt, _phase->get_ctrl(opd));
1887         }
1888         assert(opd->bottom_type()->isa_int(), "int type only");
1889         // Move non constant shift count into vector register.
1890         cnt = VectorNode::shift_count(p0, cnt, vlen, velt_basic_type(p0));
1891       }
1892       if (cnt != opd) {
1893         _igvn.register_new_node_with_optimizer(cnt);
1894         _phase->set_ctrl(cnt, _phase->get_ctrl(opd));
1895       }
1896       return cnt;
1897     }
1898     assert(!opd->is_StoreVector(), "such vector is not expected here");
1899     // Convert scalar input to vector with the same number of elements as
1900     // p0's vector. Use p0's type because size of operand's container in
1901     // vector should match p0's size regardless operand's size.
1902     const Type* p0_t = velt_type(p0);
1903     VectorNode* vn = VectorNode::scalar2vector(opd, vlen, p0_t);
1904 
1905     _igvn.register_new_node_with_optimizer(vn);
1906     _phase->set_ctrl(vn, _phase->get_ctrl(opd));
1907 #ifdef ASSERT
1908     if (TraceNewVectors) {
1909       tty->print("new Vector node: ");
1910       vn->dump();
1911     }
1912 #endif
1913     return vn;
1914   }
1915 
1916   // Insert pack operation
1917   BasicType bt = velt_basic_type(p0);
1918   PackNode* pk = PackNode::make(opd, vlen, bt);
1919   DEBUG_ONLY( const BasicType opd_bt = opd->bottom_type()->basic_type(); )
1920 
1921   for (uint i = 1; i < vlen; i++) {
1922     Node* pi = p->at(i);
1923     Node* in = pi->in(opd_idx);
1924     assert(my_pack(in) == NULL, "Should already have been unpacked");
1925     assert(opd_bt == in->bottom_type()->basic_type(), "all same type");
1926     pk->add_opd(in);
1927   }
1928   _igvn.register_new_node_with_optimizer(pk);
1929   _phase->set_ctrl(pk, _phase->get_ctrl(opd));
1930 #ifdef ASSERT
1931   if (TraceNewVectors) {
1932     tty->print("new Vector node: ");
1933     pk->dump();
1934   }
1935 #endif
1936   return pk;
1937 }
1938 
1939 //------------------------------insert_extracts---------------------------
1940 // If a use of pack p is not a vector use, then replace the
1941 // use with an extract operation.
1942 void SuperWord::insert_extracts(Node_List* p) {
1943   if (p->at(0)->is_Store()) return;
1944   assert(_n_idx_list.is_empty(), "empty (node,index) list");
1945 
1946   // Inspect each use of each pack member.  For each use that is
1947   // not a vector use, replace the use with an extract operation.
1948 
1949   for (uint i = 0; i < p->size(); i++) {
1950     Node* def = p->at(i);
1951     for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
1952       Node* use = def->fast_out(j);
1953       for (uint k = 0; k < use->req(); k++) {
1954         Node* n = use->in(k);
1955         if (def == n) {
1956           if (!is_vector_use(use, k)) {
1957             _n_idx_list.push(use, k);
1958           }
1959         }
1960       }
1961     }
1962   }
1963 
1964   while (_n_idx_list.is_nonempty()) {
1965     Node* use = _n_idx_list.node();
1966     int   idx = _n_idx_list.index();
1967     _n_idx_list.pop();
1968     Node* def = use->in(idx);
1969 
1970     if (def->is_reduction()) continue;
1971 
1972     // Insert extract operation
1973     _igvn.hash_delete(def);
1974     int def_pos = alignment(def) / data_size(def);
1975 
1976     Node* ex = ExtractNode::make(def, def_pos, velt_basic_type(def));
1977     _igvn.register_new_node_with_optimizer(ex);
1978     _phase->set_ctrl(ex, _phase->get_ctrl(def));
1979     _igvn.replace_input_of(use, idx, ex);
1980     _igvn._worklist.push(def);
1981 
1982     bb_insert_after(ex, bb_idx(def));
1983     set_velt_type(ex, velt_type(def));
1984   }
1985 }
1986 
1987 //------------------------------is_vector_use---------------------------
1988 // Is use->in(u_idx) a vector use?
1989 bool SuperWord::is_vector_use(Node* use, int u_idx) {
1990   Node_List* u_pk = my_pack(use);
1991   if (u_pk == NULL) return false;
1992   if (use->is_reduction()) return true;
1993   Node* def = use->in(u_idx);
1994   Node_List* d_pk = my_pack(def);
1995   if (d_pk == NULL) {
1996     // check for scalar promotion
1997     Node* n = u_pk->at(0)->in(u_idx);
1998     for (uint i = 1; i < u_pk->size(); i++) {
1999       if (u_pk->at(i)->in(u_idx) != n) return false;
2000     }
2001     return true;
2002   }
2003   if (u_pk->size() != d_pk->size())
2004     return false;
2005   for (uint i = 0; i < u_pk->size(); i++) {
2006     Node* ui = u_pk->at(i);
2007     Node* di = d_pk->at(i);
2008     if (ui->in(u_idx) != di || alignment(ui) != alignment(di))
2009       return false;
2010   }
2011   return true;
2012 }
2013 
2014 //------------------------------construct_bb---------------------------
2015 // Construct reverse postorder list of block members
2016 bool SuperWord::construct_bb() {
2017   Node* entry = bb();
2018 
2019   assert(_stk.length() == 0,            "stk is empty");
2020   assert(_block.length() == 0,          "block is empty");
2021   assert(_data_entry.length() == 0,     "data_entry is empty");
2022   assert(_mem_slice_head.length() == 0, "mem_slice_head is empty");
2023   assert(_mem_slice_tail.length() == 0, "mem_slice_tail is empty");
2024 
2025   // Find non-control nodes with no inputs from within block,
2026   // create a temporary map from node _idx to bb_idx for use
2027   // by the visited and post_visited sets,
2028   // and count number of nodes in block.
2029   int bb_ct = 0;
2030   for (uint i = 0; i < lpt()->_body.size(); i++) {
2031     Node *n = lpt()->_body.at(i);
2032     set_bb_idx(n, i); // Create a temporary map
2033     if (in_bb(n)) {
2034       if (n->is_LoadStore() || n->is_MergeMem() ||
2035           (n->is_Proj() && !n->as_Proj()->is_CFG())) {
2036         // Bailout if the loop has LoadStore, MergeMem or data Proj
2037         // nodes. Superword optimization does not work with them.
2038         return false;
2039       }
2040       bb_ct++;
2041       if (!n->is_CFG()) {
2042         bool found = false;
2043         for (uint j = 0; j < n->req(); j++) {
2044           Node* def = n->in(j);
2045           if (def && in_bb(def)) {
2046             found = true;
2047             break;
2048           }
2049         }
2050         if (!found) {
2051           assert(n != entry, "can't be entry");
2052           _data_entry.push(n);
2053         }
2054       }
2055     }
2056   }
2057 
2058   // Find memory slices (head and tail)
2059   for (DUIterator_Fast imax, i = lp()->fast_outs(imax); i < imax; i++) {
2060     Node *n = lp()->fast_out(i);
2061     if (in_bb(n) && (n->is_Phi() && n->bottom_type() == Type::MEMORY)) {
2062       Node* n_tail  = n->in(LoopNode::LoopBackControl);
2063       if (n_tail != n->in(LoopNode::EntryControl)) {
2064         if (!n_tail->is_Mem()) {
2065           assert(n_tail->is_Mem(), err_msg_res("unexpected node for memory slice: %s", n_tail->Name()));
2066           return false; // Bailout
2067         }
2068         _mem_slice_head.push(n);
2069         _mem_slice_tail.push(n_tail);
2070       }
2071     }
2072   }
2073 
2074   // Create an RPO list of nodes in block
2075 
2076   visited_clear();
2077   post_visited_clear();
2078 
2079   // Push all non-control nodes with no inputs from within block, then control entry
2080   for (int j = 0; j < _data_entry.length(); j++) {
2081     Node* n = _data_entry.at(j);
2082     visited_set(n);
2083     _stk.push(n);
2084   }
2085   visited_set(entry);
2086   _stk.push(entry);
2087 
2088   // Do a depth first walk over out edges
2089   int rpo_idx = bb_ct - 1;
2090   int size;
2091   int reduction_uses = 0;
2092   while ((size = _stk.length()) > 0) {
2093     Node* n = _stk.top(); // Leave node on stack
2094     if (!visited_test_set(n)) {
2095       // forward arc in graph
2096     } else if (!post_visited_test(n)) {
2097       // cross or back arc
2098       for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2099         Node *use = n->fast_out(i);
2100         if (in_bb(use) && !visited_test(use) &&
2101             // Don't go around backedge
2102             (!use->is_Phi() || n == entry)) {
2103           if (use->is_reduction()) {
2104             // First see if we can map the reduction on the given system we are on, then
2105             // make a data entry operation for each reduction we see.
2106             BasicType bt = use->bottom_type()->basic_type();
2107             if (ReductionNode::implemented(use->Opcode(), Matcher::min_vector_size(bt), bt)) {
2108               reduction_uses++;
2109             }
2110           }
2111           _stk.push(use);
2112         }
2113       }
2114       if (_stk.length() == size) {
2115         // There were no additional uses, post visit node now
2116         _stk.pop(); // Remove node from stack
2117         assert(rpo_idx >= 0, "");
2118         _block.at_put_grow(rpo_idx, n);
2119         rpo_idx--;
2120         post_visited_set(n);
2121         assert(rpo_idx >= 0 || _stk.is_empty(), "");
2122       }
2123     } else {
2124       _stk.pop(); // Remove post-visited node from stack
2125     }
2126   }
2127 
2128   // Create real map of block indices for nodes
2129   for (int j = 0; j < _block.length(); j++) {
2130     Node* n = _block.at(j);
2131     set_bb_idx(n, j);
2132   }
2133 
2134   // Ensure extra info is allocated.
2135   initialize_bb();
2136 
2137 #ifndef PRODUCT
2138   if (TraceSuperWord) {
2139     print_bb();
2140     tty->print_cr("\ndata entry nodes: %s", _data_entry.length() > 0 ? "" : "NONE");
2141     for (int m = 0; m < _data_entry.length(); m++) {
2142       tty->print("%3d ", m);
2143       _data_entry.at(m)->dump();
2144     }
2145     tty->print_cr("\nmemory slices: %s", _mem_slice_head.length() > 0 ? "" : "NONE");
2146     for (int m = 0; m < _mem_slice_head.length(); m++) {
2147       tty->print("%3d ", m); _mem_slice_head.at(m)->dump();
2148       tty->print("    ");    _mem_slice_tail.at(m)->dump();
2149     }
2150   }
2151 #endif
2152   assert(rpo_idx == -1 && bb_ct == _block.length(), "all block members found");
2153   return (_mem_slice_head.length() > 0) || (reduction_uses > 0) || (_data_entry.length() > 0);
2154 }
2155 
2156 //------------------------------initialize_bb---------------------------
2157 // Initialize per node info
2158 void SuperWord::initialize_bb() {
2159   Node* last = _block.at(_block.length() - 1);
2160   grow_node_info(bb_idx(last));
2161 }
2162 
2163 //------------------------------bb_insert_after---------------------------
2164 // Insert n into block after pos
2165 void SuperWord::bb_insert_after(Node* n, int pos) {
2166   int n_pos = pos + 1;
2167   // Make room
2168   for (int i = _block.length() - 1; i >= n_pos; i--) {
2169     _block.at_put_grow(i+1, _block.at(i));
2170   }
2171   for (int j = _node_info.length() - 1; j >= n_pos; j--) {
2172     _node_info.at_put_grow(j+1, _node_info.at(j));
2173   }
2174   // Set value
2175   _block.at_put_grow(n_pos, n);
2176   _node_info.at_put_grow(n_pos, SWNodeInfo::initial);
2177   // Adjust map from node->_idx to _block index
2178   for (int i = n_pos; i < _block.length(); i++) {
2179     set_bb_idx(_block.at(i), i);
2180   }
2181 }
2182 
2183 //------------------------------compute_max_depth---------------------------
2184 // Compute max depth for expressions from beginning of block
2185 // Use to prune search paths during test for independence.
2186 void SuperWord::compute_max_depth() {
2187   int ct = 0;
2188   bool again;
2189   do {
2190     again = false;
2191     for (int i = 0; i < _block.length(); i++) {
2192       Node* n = _block.at(i);
2193       if (!n->is_Phi()) {
2194         int d_orig = depth(n);
2195         int d_in   = 0;
2196         for (DepPreds preds(n, _dg); !preds.done(); preds.next()) {
2197           Node* pred = preds.current();
2198           if (in_bb(pred)) {
2199             d_in = MAX2(d_in, depth(pred));
2200           }
2201         }
2202         if (d_in + 1 != d_orig) {
2203           set_depth(n, d_in + 1);
2204           again = true;
2205         }
2206       }
2207     }
2208     ct++;
2209   } while (again);
2210 #ifndef PRODUCT
2211   if (TraceSuperWord && Verbose)
2212     tty->print_cr("compute_max_depth iterated: %d times", ct);
2213 #endif
2214 }
2215 
2216 //-------------------------compute_vector_element_type-----------------------
2217 // Compute necessary vector element type for expressions
2218 // This propagates backwards a narrower integer type when the
2219 // upper bits of the value are not needed.
2220 // Example:  char a,b,c;  a = b + c;
2221 // Normally the type of the add is integer, but for packed character
2222 // operations the type of the add needs to be char.
2223 void SuperWord::compute_vector_element_type() {
2224 #ifndef PRODUCT
2225   if (TraceSuperWord && Verbose)
2226     tty->print_cr("\ncompute_velt_type:");
2227 #endif
2228 
2229   // Initial type
2230   for (int i = 0; i < _block.length(); i++) {
2231     Node* n = _block.at(i);
2232     set_velt_type(n, container_type(n));
2233   }
2234 
2235   // Propagate integer narrowed type backwards through operations
2236   // that don't depend on higher order bits
2237   for (int i = _block.length() - 1; i >= 0; i--) {
2238     Node* n = _block.at(i);
2239     // Only integer types need be examined
2240     const Type* vtn = velt_type(n);
2241     if (vtn->basic_type() == T_INT) {
2242       uint start, end;
2243       VectorNode::vector_operands(n, &start, &end);
2244 
2245       for (uint j = start; j < end; j++) {
2246         Node* in  = n->in(j);
2247         // Don't propagate through a memory
2248         if (!in->is_Mem() && in_bb(in) && velt_type(in)->basic_type() == T_INT &&
2249             data_size(n) < data_size(in)) {
2250           bool same_type = true;
2251           for (DUIterator_Fast kmax, k = in->fast_outs(kmax); k < kmax; k++) {
2252             Node *use = in->fast_out(k);
2253             if (!in_bb(use) || !same_velt_type(use, n)) {
2254               same_type = false;
2255               break;
2256             }
2257           }
2258           if (same_type) {
2259             // For right shifts of small integer types (bool, byte, char, short)
2260             // we need precise information about sign-ness. Only Load nodes have
2261             // this information because Store nodes are the same for signed and
2262             // unsigned values. And any arithmetic operation after a load may
2263             // expand a value to signed Int so such right shifts can't be used
2264             // because vector elements do not have upper bits of Int.
2265             const Type* vt = vtn;
2266             if (VectorNode::is_shift(in)) {
2267               Node* load = in->in(1);
2268               if (load->is_Load() && in_bb(load) && (velt_type(load)->basic_type() == T_INT)) {
2269                 vt = velt_type(load);
2270               } else if (in->Opcode() != Op_LShiftI) {
2271                 // Widen type to Int to avoid creation of right shift vector
2272                 // (align + data_size(s1) check in stmts_can_pack() will fail).
2273                 // Note, left shifts work regardless type.
2274                 vt = TypeInt::INT;
2275               }
2276             }
2277             set_velt_type(in, vt);
2278           }
2279         }
2280       }
2281     }
2282   }
2283 #ifndef PRODUCT
2284   if (TraceSuperWord && Verbose) {
2285     for (int i = 0; i < _block.length(); i++) {
2286       Node* n = _block.at(i);
2287       velt_type(n)->dump();
2288       tty->print("\t");
2289       n->dump();
2290     }
2291   }
2292 #endif
2293 }
2294 
2295 //------------------------------memory_alignment---------------------------
2296 // Alignment within a vector memory reference
2297 int SuperWord::memory_alignment(MemNode* s, int iv_adjust) {
2298   SWPointer p(s, this, NULL, false);
2299   if (!p.valid()) {
2300     return bottom_align;
2301   }
2302   int vw = vector_width_in_bytes(s);
2303   if (vw < 2) {
2304     return bottom_align; // No vectors for this type
2305   }
2306   int offset  = p.offset_in_bytes();
2307   offset     += iv_adjust*p.memory_size();
2308   int off_rem = offset % vw;
2309   int off_mod = off_rem >= 0 ? off_rem : off_rem + vw;
2310   return off_mod;
2311 }
2312 
2313 //---------------------------container_type---------------------------
2314 // Smallest type containing range of values
2315 const Type* SuperWord::container_type(Node* n) {
2316   if (n->is_Mem()) {
2317     BasicType bt = n->as_Mem()->memory_type();
2318     if (n->is_Store() && (bt == T_CHAR)) {
2319       // Use T_SHORT type instead of T_CHAR for stored values because any
2320       // preceding arithmetic operation extends values to signed Int.
2321       bt = T_SHORT;
2322     }
2323     if (n->Opcode() == Op_LoadUB) {
2324       // Adjust type for unsigned byte loads, it is important for right shifts.
2325       // T_BOOLEAN is used because there is no basic type representing type
2326       // TypeInt::UBYTE. Use of T_BOOLEAN for vectors is fine because only
2327       // size (one byte) and sign is important.
2328       bt = T_BOOLEAN;
2329     }
2330     return Type::get_const_basic_type(bt);
2331   }
2332   const Type* t = _igvn.type(n);
2333   if (t->basic_type() == T_INT) {
2334     // A narrow type of arithmetic operations will be determined by
2335     // propagating the type of memory operations.
2336     return TypeInt::INT;
2337   }
2338   return t;
2339 }
2340 
2341 bool SuperWord::same_velt_type(Node* n1, Node* n2) {
2342   const Type* vt1 = velt_type(n1);
2343   const Type* vt2 = velt_type(n2);
2344   if (vt1->basic_type() == T_INT && vt2->basic_type() == T_INT) {
2345     // Compare vectors element sizes for integer types.
2346     return data_size(n1) == data_size(n2);
2347   }
2348   return vt1 == vt2;
2349 }
2350 
2351 //------------------------------in_packset---------------------------
2352 // Are s1 and s2 in a pack pair and ordered as s1,s2?
2353 bool SuperWord::in_packset(Node* s1, Node* s2) {
2354   for (int i = 0; i < _packset.length(); i++) {
2355     Node_List* p = _packset.at(i);
2356     assert(p->size() == 2, "must be");
2357     if (p->at(0) == s1 && p->at(p->size()-1) == s2) {
2358       return true;
2359     }
2360   }
2361   return false;
2362 }
2363 
2364 //------------------------------in_pack---------------------------
2365 // Is s in pack p?
2366 Node_List* SuperWord::in_pack(Node* s, Node_List* p) {
2367   for (uint i = 0; i < p->size(); i++) {
2368     if (p->at(i) == s) {
2369       return p;
2370     }
2371   }
2372   return NULL;
2373 }
2374 
2375 //------------------------------remove_pack_at---------------------------
2376 // Remove the pack at position pos in the packset
2377 void SuperWord::remove_pack_at(int pos) {
2378   Node_List* p = _packset.at(pos);
2379   for (uint i = 0; i < p->size(); i++) {
2380     Node* s = p->at(i);
2381     set_my_pack(s, NULL);
2382   }
2383   _packset.remove_at(pos);
2384 }
2385 
2386 void SuperWord::packset_sort(int n) {
2387   // simple bubble sort so that we capitalize with O(n) when its already sorted
2388   while (n != 0) {
2389     bool swapped = false;
2390     for (int i = 1; i < n; i++) {
2391       Node_List* q_low = _packset.at(i-1);
2392       Node_List* q_i = _packset.at(i);
2393 
2394       // only swap when we find something to swap
2395       if (alignment(q_low->at(0)) > alignment(q_i->at(0))) {
2396         Node_List* t = q_i;
2397         *(_packset.adr_at(i)) = q_low;
2398         *(_packset.adr_at(i-1)) = q_i;
2399         swapped = true;
2400       }
2401     }
2402     if (swapped == false) break;
2403     n--;
2404   }
2405 }
2406 
2407 //------------------------------executed_first---------------------------
2408 // Return the node executed first in pack p.  Uses the RPO block list
2409 // to determine order.
2410 Node* SuperWord::executed_first(Node_List* p) {
2411   Node* n = p->at(0);
2412   int n_rpo = bb_idx(n);
2413   for (uint i = 1; i < p->size(); i++) {
2414     Node* s = p->at(i);
2415     int s_rpo = bb_idx(s);
2416     if (s_rpo < n_rpo) {
2417       n = s;
2418       n_rpo = s_rpo;
2419     }
2420   }
2421   return n;
2422 }
2423 
2424 //------------------------------executed_last---------------------------
2425 // Return the node executed last in pack p.
2426 Node* SuperWord::executed_last(Node_List* p) {
2427   Node* n = p->at(0);
2428   int n_rpo = bb_idx(n);
2429   for (uint i = 1; i < p->size(); i++) {
2430     Node* s = p->at(i);
2431     int s_rpo = bb_idx(s);
2432     if (s_rpo > n_rpo) {
2433       n = s;
2434       n_rpo = s_rpo;
2435     }
2436   }
2437   return n;
2438 }
2439 
2440 LoadNode::ControlDependency SuperWord::control_dependency(Node_List* p) {
2441   LoadNode::ControlDependency dep = LoadNode::DependsOnlyOnTest;
2442   for (uint i = 0; i < p->size(); i++) {
2443     Node* n = p->at(i);
2444     assert(n->is_Load(), "only meaningful for loads");
2445     if (!n->depends_only_on_test()) {
2446       dep = LoadNode::Pinned;
2447     }
2448   }
2449   return dep;
2450 }
2451 
2452 
2453 //----------------------------align_initial_loop_index---------------------------
2454 // Adjust pre-loop limit so that in main loop, a load/store reference
2455 // to align_to_ref will be a position zero in the vector.
2456 //   (iv + k) mod vector_align == 0
2457 void SuperWord::align_initial_loop_index(MemNode* align_to_ref) {
2458   CountedLoopNode *main_head = lp()->as_CountedLoop();
2459   assert(main_head->is_main_loop(), "");
2460   CountedLoopEndNode* pre_end = get_pre_loop_end(main_head);
2461   assert(pre_end != NULL, "we must have a correct pre-loop");
2462   Node *pre_opaq1 = pre_end->limit();
2463   assert(pre_opaq1->Opcode() == Op_Opaque1, "");
2464   Opaque1Node *pre_opaq = (Opaque1Node*)pre_opaq1;
2465   Node *lim0 = pre_opaq->in(1);
2466 
2467   // Where we put new limit calculations
2468   Node *pre_ctrl = pre_end->loopnode()->in(LoopNode::EntryControl);
2469 
2470   // Ensure the original loop limit is available from the
2471   // pre-loop Opaque1 node.
2472   Node *orig_limit = pre_opaq->original_loop_limit();
2473   assert(orig_limit != NULL && _igvn.type(orig_limit) != Type::TOP, "");
2474 
2475   SWPointer align_to_ref_p(align_to_ref, this, NULL, false);
2476   assert(align_to_ref_p.valid(), "sanity");
2477 
2478   // Given:
2479   //     lim0 == original pre loop limit
2480   //     V == v_align (power of 2)
2481   //     invar == extra invariant piece of the address expression
2482   //     e == offset [ +/- invar ]
2483   //
2484   // When reassociating expressions involving '%' the basic rules are:
2485   //     (a - b) % k == 0   =>  a % k == b % k
2486   // and:
2487   //     (a + b) % k == 0   =>  a % k == (k - b) % k
2488   //
2489   // For stride > 0 && scale > 0,
2490   //   Derive the new pre-loop limit "lim" such that the two constraints:
2491   //     (1) lim = lim0 + N           (where N is some positive integer < V)
2492   //     (2) (e + lim) % V == 0
2493   //   are true.
2494   //
2495   //   Substituting (1) into (2),
2496   //     (e + lim0 + N) % V == 0
2497   //   solve for N:
2498   //     N = (V - (e + lim0)) % V
2499   //   substitute back into (1), so that new limit
2500   //     lim = lim0 + (V - (e + lim0)) % V
2501   //
2502   // For stride > 0 && scale < 0
2503   //   Constraints:
2504   //     lim = lim0 + N
2505   //     (e - lim) % V == 0
2506   //   Solving for lim:
2507   //     (e - lim0 - N) % V == 0
2508   //     N = (e - lim0) % V
2509   //     lim = lim0 + (e - lim0) % V
2510   //
2511   // For stride < 0 && scale > 0
2512   //   Constraints:
2513   //     lim = lim0 - N
2514   //     (e + lim) % V == 0
2515   //   Solving for lim:
2516   //     (e + lim0 - N) % V == 0
2517   //     N = (e + lim0) % V
2518   //     lim = lim0 - (e + lim0) % V
2519   //
2520   // For stride < 0 && scale < 0
2521   //   Constraints:
2522   //     lim = lim0 - N
2523   //     (e - lim) % V == 0
2524   //   Solving for lim:
2525   //     (e - lim0 + N) % V == 0
2526   //     N = (V - (e - lim0)) % V
2527   //     lim = lim0 - (V - (e - lim0)) % V
2528 
2529   int vw = vector_width_in_bytes(align_to_ref);
2530   int stride   = iv_stride();
2531   int scale    = align_to_ref_p.scale_in_bytes();
2532   int elt_size = align_to_ref_p.memory_size();
2533   int v_align  = vw / elt_size;
2534   assert(v_align > 1, "sanity");
2535   int offset   = align_to_ref_p.offset_in_bytes() / elt_size;
2536   Node *offsn  = _igvn.intcon(offset);
2537 
2538   Node *e = offsn;
2539   if (align_to_ref_p.invar() != NULL) {
2540     // incorporate any extra invariant piece producing (offset +/- invar) >>> log2(elt)
2541     Node* log2_elt = _igvn.intcon(exact_log2(elt_size));
2542     Node* aref     = new URShiftINode(align_to_ref_p.invar(), log2_elt);
2543     _igvn.register_new_node_with_optimizer(aref);
2544     _phase->set_ctrl(aref, pre_ctrl);
2545     if (align_to_ref_p.negate_invar()) {
2546       e = new SubINode(e, aref);
2547     } else {
2548       e = new AddINode(e, aref);
2549     }
2550     _igvn.register_new_node_with_optimizer(e);
2551     _phase->set_ctrl(e, pre_ctrl);
2552   }
2553   if (vw > ObjectAlignmentInBytes) {
2554     // incorporate base e +/- base && Mask >>> log2(elt)
2555     Node* xbase = new CastP2XNode(NULL, align_to_ref_p.base());
2556     _igvn.register_new_node_with_optimizer(xbase);
2557 #ifdef _LP64
2558     xbase  = new ConvL2INode(xbase);
2559     _igvn.register_new_node_with_optimizer(xbase);
2560 #endif
2561     Node* mask = _igvn.intcon(vw-1);
2562     Node* masked_xbase  = new AndINode(xbase, mask);
2563     _igvn.register_new_node_with_optimizer(masked_xbase);
2564     Node* log2_elt = _igvn.intcon(exact_log2(elt_size));
2565     Node* bref     = new URShiftINode(masked_xbase, log2_elt);
2566     _igvn.register_new_node_with_optimizer(bref);
2567     _phase->set_ctrl(bref, pre_ctrl);
2568     e = new AddINode(e, bref);
2569     _igvn.register_new_node_with_optimizer(e);
2570     _phase->set_ctrl(e, pre_ctrl);
2571   }
2572 
2573   // compute e +/- lim0
2574   if (scale < 0) {
2575     e = new SubINode(e, lim0);
2576   } else {
2577     e = new AddINode(e, lim0);
2578   }
2579   _igvn.register_new_node_with_optimizer(e);
2580   _phase->set_ctrl(e, pre_ctrl);
2581 
2582   if (stride * scale > 0) {
2583     // compute V - (e +/- lim0)
2584     Node* va  = _igvn.intcon(v_align);
2585     e = new SubINode(va, e);
2586     _igvn.register_new_node_with_optimizer(e);
2587     _phase->set_ctrl(e, pre_ctrl);
2588   }
2589   // compute N = (exp) % V
2590   Node* va_msk = _igvn.intcon(v_align - 1);
2591   Node* N = new AndINode(e, va_msk);
2592   _igvn.register_new_node_with_optimizer(N);
2593   _phase->set_ctrl(N, pre_ctrl);
2594 
2595   //   substitute back into (1), so that new limit
2596   //     lim = lim0 + N
2597   Node* lim;
2598   if (stride < 0) {
2599     lim = new SubINode(lim0, N);
2600   } else {
2601     lim = new AddINode(lim0, N);
2602   }
2603   _igvn.register_new_node_with_optimizer(lim);
2604   _phase->set_ctrl(lim, pre_ctrl);
2605   Node* constrained =
2606     (stride > 0) ? (Node*) new MinINode(lim, orig_limit)
2607                  : (Node*) new MaxINode(lim, orig_limit);
2608   _igvn.register_new_node_with_optimizer(constrained);
2609   _phase->set_ctrl(constrained, pre_ctrl);
2610   _igvn.hash_delete(pre_opaq);
2611   pre_opaq->set_req(1, constrained);
2612 }
2613 
2614 //----------------------------get_pre_loop_end---------------------------
2615 // Find pre loop end from main loop.  Returns null if none.
2616 CountedLoopEndNode* SuperWord::get_pre_loop_end(CountedLoopNode *cl) {
2617   Node *ctrl = cl->in(LoopNode::EntryControl);
2618   if (!ctrl->is_IfTrue() && !ctrl->is_IfFalse()) return NULL;
2619   Node *iffm = ctrl->in(0);
2620   if (!iffm->is_If()) return NULL;
2621   Node *p_f = iffm->in(0);
2622   if (!p_f->is_IfFalse()) return NULL;
2623   if (!p_f->in(0)->is_CountedLoopEnd()) return NULL;
2624   CountedLoopEndNode *pre_end = p_f->in(0)->as_CountedLoopEnd();
2625   CountedLoopNode* loop_node = pre_end->loopnode();
2626   if (loop_node == NULL || !loop_node->is_pre_loop()) return NULL;
2627   return pre_end;
2628 }
2629 
2630 
2631 //------------------------------init---------------------------
2632 void SuperWord::init() {
2633   _dg.init();
2634   _packset.clear();
2635   _disjoint_ptrs.clear();
2636   _block.clear();
2637   _data_entry.clear();
2638   _mem_slice_head.clear();
2639   _mem_slice_tail.clear();
2640   _iteration_first.clear();
2641   _iteration_last.clear();
2642   _node_info.clear();
2643   _align_to_ref = NULL;
2644   _lpt = NULL;
2645   _lp = NULL;
2646   _bb = NULL;
2647   _iv = NULL;
2648   _race_possible = 0;
2649   _early_return = false;
2650   _num_work_vecs = 0;
2651   _num_reductions = 0;
2652 }
2653 
2654 //------------------------------restart---------------------------
2655 void SuperWord::restart() {
2656   _dg.init();
2657   _packset.clear();
2658   _disjoint_ptrs.clear();
2659   _block.clear();
2660   _data_entry.clear();
2661   _mem_slice_head.clear();
2662   _mem_slice_tail.clear();
2663   _node_info.clear();
2664 }
2665 
2666 //------------------------------print_packset---------------------------
2667 void SuperWord::print_packset() {
2668 #ifndef PRODUCT
2669   tty->print_cr("packset");
2670   for (int i = 0; i < _packset.length(); i++) {
2671     tty->print_cr("Pack: %d", i);
2672     Node_List* p = _packset.at(i);
2673     print_pack(p);
2674   }
2675 #endif
2676 }
2677 
2678 //------------------------------print_pack---------------------------
2679 void SuperWord::print_pack(Node_List* p) {
2680   for (uint i = 0; i < p->size(); i++) {
2681     print_stmt(p->at(i));
2682   }
2683 }
2684 
2685 //------------------------------print_bb---------------------------
2686 void SuperWord::print_bb() {
2687 #ifndef PRODUCT
2688   tty->print_cr("\nBlock");
2689   for (int i = 0; i < _block.length(); i++) {
2690     Node* n = _block.at(i);
2691     tty->print("%d ", i);
2692     if (n) {
2693       n->dump();
2694     }
2695   }
2696 #endif
2697 }
2698 
2699 //------------------------------print_stmt---------------------------
2700 void SuperWord::print_stmt(Node* s) {
2701 #ifndef PRODUCT
2702   tty->print(" align: %d \t", alignment(s));
2703   s->dump();
2704 #endif
2705 }
2706 
2707 //------------------------------blank---------------------------
2708 char* SuperWord::blank(uint depth) {
2709   static char blanks[101];
2710   assert(depth < 101, "too deep");
2711   for (uint i = 0; i < depth; i++) blanks[i] = ' ';
2712   blanks[depth] = '\0';
2713   return blanks;
2714 }
2715 
2716 
2717 //==============================SWPointer===========================
2718 
2719 //----------------------------SWPointer------------------------
2720 SWPointer::SWPointer(MemNode* mem, SuperWord* slp, Node_Stack *nstack, bool analyze_only) :
2721   _mem(mem), _slp(slp),  _base(NULL),  _adr(NULL),
2722   _scale(0), _offset(0), _invar(NULL), _negate_invar(false),
2723   _nstack(nstack), _analyze_only(analyze_only),
2724   _stack_idx(0) {
2725 
2726   Node* adr = mem->in(MemNode::Address);
2727   if (!adr->is_AddP()) {
2728     assert(!valid(), "too complex");
2729     return;
2730   }
2731   // Match AddP(base, AddP(ptr, k*iv [+ invariant]), constant)
2732   Node* base = adr->in(AddPNode::Base);
2733   // The base address should be loop invariant
2734   if (!invariant(base)) {
2735     assert(!valid(), "base address is loop variant");
2736     return;
2737   }
2738   //unsafe reference could not be aligned appropriately without runtime checking
2739   if (base == NULL || base->bottom_type() == Type::TOP) {
2740     assert(!valid(), "unsafe access");
2741     return;
2742   }
2743   for (int i = 0; i < 3; i++) {
2744     if (!scaled_iv_plus_offset(adr->in(AddPNode::Offset))) {
2745       assert(!valid(), "too complex");
2746       return;
2747     }
2748     adr = adr->in(AddPNode::Address);
2749     if (base == adr || !adr->is_AddP()) {
2750       break; // stop looking at addp's
2751     }
2752   }
2753   _base = base;
2754   _adr  = adr;
2755   assert(valid(), "Usable");
2756 }
2757 
2758 // Following is used to create a temporary object during
2759 // the pattern match of an address expression.
2760 SWPointer::SWPointer(SWPointer* p) :
2761   _mem(p->_mem), _slp(p->_slp),  _base(NULL),  _adr(NULL),
2762   _scale(0), _offset(0), _invar(NULL), _negate_invar(false),
2763   _nstack(p->_nstack), _analyze_only(p->_analyze_only),
2764   _stack_idx(p->_stack_idx) {}
2765 
2766 //------------------------scaled_iv_plus_offset--------------------
2767 // Match: k*iv + offset
2768 // where: k is a constant that maybe zero, and
2769 //        offset is (k2 [+/- invariant]) where k2 maybe zero and invariant is optional
2770 bool SWPointer::scaled_iv_plus_offset(Node* n) {
2771   if (scaled_iv(n)) {
2772     return true;
2773   }
2774   if (offset_plus_k(n)) {
2775     return true;
2776   }
2777   int opc = n->Opcode();
2778   if (opc == Op_AddI) {
2779     if (scaled_iv(n->in(1)) && offset_plus_k(n->in(2))) {
2780       return true;
2781     }
2782     if (scaled_iv(n->in(2)) && offset_plus_k(n->in(1))) {
2783       return true;
2784     }
2785   } else if (opc == Op_SubI) {
2786     if (scaled_iv(n->in(1)) && offset_plus_k(n->in(2), true)) {
2787       return true;
2788     }
2789     if (scaled_iv(n->in(2)) && offset_plus_k(n->in(1))) {
2790       _scale *= -1;
2791       return true;
2792     }
2793   }
2794   return false;
2795 }
2796 
2797 //----------------------------scaled_iv------------------------
2798 // Match: k*iv where k is a constant that's not zero
2799 bool SWPointer::scaled_iv(Node* n) {
2800   if (_scale != 0) {
2801     return false;  // already found a scale
2802   }
2803   if (n == iv()) {
2804     _scale = 1;
2805     return true;
2806   }
2807   if (_analyze_only && (invariant(n) == false)) {
2808     _nstack->push(n, _stack_idx++);
2809   }
2810   int opc = n->Opcode();
2811   if (opc == Op_MulI) {
2812     if (n->in(1) == iv() && n->in(2)->is_Con()) {
2813       _scale = n->in(2)->get_int();
2814       return true;
2815     } else if (n->in(2) == iv() && n->in(1)->is_Con()) {
2816       _scale = n->in(1)->get_int();
2817       return true;
2818     }
2819   } else if (opc == Op_LShiftI) {
2820     if (n->in(1) == iv() && n->in(2)->is_Con()) {
2821       _scale = 1 << n->in(2)->get_int();
2822       return true;
2823     }
2824   } else if (opc == Op_ConvI2L) {
2825     if (scaled_iv_plus_offset(n->in(1))) {
2826       return true;
2827     }
2828   } else if (opc == Op_LShiftL) {
2829     if (!has_iv() && _invar == NULL) {
2830       // Need to preserve the current _offset value, so
2831       // create a temporary object for this expression subtree.
2832       // Hacky, so should re-engineer the address pattern match.
2833       SWPointer tmp(this);
2834       if (tmp.scaled_iv_plus_offset(n->in(1))) {
2835         if (tmp._invar == NULL) {
2836           int mult = 1 << n->in(2)->get_int();
2837           _scale   = tmp._scale  * mult;
2838           _offset += tmp._offset * mult;
2839           return true;
2840         }
2841       }
2842     }
2843   }
2844   return false;
2845 }
2846 
2847 //----------------------------offset_plus_k------------------------
2848 // Match: offset is (k [+/- invariant])
2849 // where k maybe zero and invariant is optional, but not both.
2850 bool SWPointer::offset_plus_k(Node* n, bool negate) {
2851   int opc = n->Opcode();
2852   if (opc == Op_ConI) {
2853     _offset += negate ? -(n->get_int()) : n->get_int();
2854     return true;
2855   } else if (opc == Op_ConL) {
2856     // Okay if value fits into an int
2857     const TypeLong* t = n->find_long_type();
2858     if (t->higher_equal(TypeLong::INT)) {
2859       jlong loff = n->get_long();
2860       jint  off  = (jint)loff;
2861       _offset += negate ? -off : loff;
2862       return true;
2863     }
2864     return false;
2865   }
2866   if (_invar != NULL) return false; // already have an invariant
2867   if (_analyze_only && (invariant(n) == false)) {
2868     _nstack->push(n, _stack_idx++);
2869   }
2870   if (opc == Op_AddI) {
2871     if (n->in(2)->is_Con() && invariant(n->in(1))) {
2872       _negate_invar = negate;
2873       _invar = n->in(1);
2874       _offset += negate ? -(n->in(2)->get_int()) : n->in(2)->get_int();
2875       return true;
2876     } else if (n->in(1)->is_Con() && invariant(n->in(2))) {
2877       _offset += negate ? -(n->in(1)->get_int()) : n->in(1)->get_int();
2878       _negate_invar = negate;
2879       _invar = n->in(2);
2880       return true;
2881     }
2882   }
2883   if (opc == Op_SubI) {
2884     if (n->in(2)->is_Con() && invariant(n->in(1))) {
2885       _negate_invar = negate;
2886       _invar = n->in(1);
2887       _offset += !negate ? -(n->in(2)->get_int()) : n->in(2)->get_int();
2888       return true;
2889     } else if (n->in(1)->is_Con() && invariant(n->in(2))) {
2890       _offset += negate ? -(n->in(1)->get_int()) : n->in(1)->get_int();
2891       _negate_invar = !negate;
2892       _invar = n->in(2);
2893       return true;
2894     }
2895   }
2896   if (invariant(n)) {
2897     _negate_invar = negate;
2898     _invar = n;
2899     return true;
2900   }
2901   return false;
2902 }
2903 
2904 //----------------------------print------------------------
2905 void SWPointer::print() {
2906 #ifndef PRODUCT
2907   tty->print("base: %d  adr: %d  scale: %d  offset: %d  invar: %c%d\n",
2908              _base != NULL ? _base->_idx : 0,
2909              _adr  != NULL ? _adr->_idx  : 0,
2910              _scale, _offset,
2911              _negate_invar?'-':'+',
2912              _invar != NULL ? _invar->_idx : 0);
2913 #endif
2914 }
2915 
2916 // ========================= OrderedPair =====================
2917 
2918 const OrderedPair OrderedPair::initial;
2919 
2920 // ========================= SWNodeInfo =====================
2921 
2922 const SWNodeInfo SWNodeInfo::initial;
2923 
2924 
2925 // ============================ DepGraph ===========================
2926 
2927 //------------------------------make_node---------------------------
2928 // Make a new dependence graph node for an ideal node.
2929 DepMem* DepGraph::make_node(Node* node) {
2930   DepMem* m = new (_arena) DepMem(node);
2931   if (node != NULL) {
2932     assert(_map.at_grow(node->_idx) == NULL, "one init only");
2933     _map.at_put_grow(node->_idx, m);
2934   }
2935   return m;
2936 }
2937 
2938 //------------------------------make_edge---------------------------
2939 // Make a new dependence graph edge from dpred -> dsucc
2940 DepEdge* DepGraph::make_edge(DepMem* dpred, DepMem* dsucc) {
2941   DepEdge* e = new (_arena) DepEdge(dpred, dsucc, dsucc->in_head(), dpred->out_head());
2942   dpred->set_out_head(e);
2943   dsucc->set_in_head(e);
2944   return e;
2945 }
2946 
2947 // ========================== DepMem ========================
2948 
2949 //------------------------------in_cnt---------------------------
2950 int DepMem::in_cnt() {
2951   int ct = 0;
2952   for (DepEdge* e = _in_head; e != NULL; e = e->next_in()) ct++;
2953   return ct;
2954 }
2955 
2956 //------------------------------out_cnt---------------------------
2957 int DepMem::out_cnt() {
2958   int ct = 0;
2959   for (DepEdge* e = _out_head; e != NULL; e = e->next_out()) ct++;
2960   return ct;
2961 }
2962 
2963 //------------------------------print-----------------------------
2964 void DepMem::print() {
2965 #ifndef PRODUCT
2966   tty->print("  DepNode %d (", _node->_idx);
2967   for (DepEdge* p = _in_head; p != NULL; p = p->next_in()) {
2968     Node* pred = p->pred()->node();
2969     tty->print(" %d", pred != NULL ? pred->_idx : 0);
2970   }
2971   tty->print(") [");
2972   for (DepEdge* s = _out_head; s != NULL; s = s->next_out()) {
2973     Node* succ = s->succ()->node();
2974     tty->print(" %d", succ != NULL ? succ->_idx : 0);
2975   }
2976   tty->print_cr(" ]");
2977 #endif
2978 }
2979 
2980 // =========================== DepEdge =========================
2981 
2982 //------------------------------DepPreds---------------------------
2983 void DepEdge::print() {
2984 #ifndef PRODUCT
2985   tty->print_cr("DepEdge: %d [ %d ]", _pred->node()->_idx, _succ->node()->_idx);
2986 #endif
2987 }
2988 
2989 // =========================== DepPreds =========================
2990 // Iterator over predecessor edges in the dependence graph.
2991 
2992 //------------------------------DepPreds---------------------------
2993 DepPreds::DepPreds(Node* n, DepGraph& dg) {
2994   _n = n;
2995   _done = false;
2996   if (_n->is_Store() || _n->is_Load()) {
2997     _next_idx = MemNode::Address;
2998     _end_idx  = n->req();
2999     _dep_next = dg.dep(_n)->in_head();
3000   } else if (_n->is_Mem()) {
3001     _next_idx = 0;
3002     _end_idx  = 0;
3003     _dep_next = dg.dep(_n)->in_head();
3004   } else {
3005     _next_idx = 1;
3006     _end_idx  = _n->req();
3007     _dep_next = NULL;
3008   }
3009   next();
3010 }
3011 
3012 //------------------------------next---------------------------
3013 void DepPreds::next() {
3014   if (_dep_next != NULL) {
3015     _current  = _dep_next->pred()->node();
3016     _dep_next = _dep_next->next_in();
3017   } else if (_next_idx < _end_idx) {
3018     _current  = _n->in(_next_idx++);
3019   } else {
3020     _done = true;
3021   }
3022 }
3023 
3024 // =========================== DepSuccs =========================
3025 // Iterator over successor edges in the dependence graph.
3026 
3027 //------------------------------DepSuccs---------------------------
3028 DepSuccs::DepSuccs(Node* n, DepGraph& dg) {
3029   _n = n;
3030   _done = false;
3031   if (_n->is_Load()) {
3032     _next_idx = 0;
3033     _end_idx  = _n->outcnt();
3034     _dep_next = dg.dep(_n)->out_head();
3035   } else if (_n->is_Mem() || _n->is_Phi() && _n->bottom_type() == Type::MEMORY) {
3036     _next_idx = 0;
3037     _end_idx  = 0;
3038     _dep_next = dg.dep(_n)->out_head();
3039   } else {
3040     _next_idx = 0;
3041     _end_idx  = _n->outcnt();
3042     _dep_next = NULL;
3043   }
3044   next();
3045 }
3046 
3047 //-------------------------------next---------------------------
3048 void DepSuccs::next() {
3049   if (_dep_next != NULL) {
3050     _current  = _dep_next->succ()->node();
3051     _dep_next = _dep_next->next_out();
3052   } else if (_next_idx < _end_idx) {
3053     _current  = _n->raw_out(_next_idx++);
3054   } else {
3055     _done = true;
3056   }
3057 }
3058 
3059 //
3060 // --------------------------------- vectorization/simd -----------------------------------
3061 //
3062 Node*  SuperWord::find_phi_for_mem_dep(LoadNode* ld) {
3063   assert(in_bb(ld), "must be in block");
3064   if (_clone_map.gen(ld->_idx) == _ii_first) {
3065 #ifndef PRODUCT
3066     if (_vector_loop_debug) {
3067       tty->print_cr("SuperWord::find_phi_for_mem_dep _clone_map.gen(ld->_idx)=%d",
3068                     _clone_map.gen(ld->_idx));
3069     }
3070 #endif
3071     return NULL; //we think that any ld in the first gen being vectorizable
3072   }
3073 
3074   Node* mem = ld->in(MemNode::Memory);
3075   if (mem->outcnt() <= 1) {
3076     // we don't want to remove the only edge from mem node to load
3077 #ifndef PRODUCT
3078     if (_vector_loop_debug) {
3079       tty->print_cr("SuperWord::find_phi_for_mem_dep input node %d to load %d has no other outputs and edge mem->load cannot be removed",
3080                     mem->_idx, ld->_idx);
3081       ld->dump();
3082       mem->dump();
3083     }
3084 #endif
3085     return NULL;
3086   }
3087   if (!in_bb(mem) || _clone_map.gen(mem->_idx) == _clone_map.gen(ld->_idx)) {
3088 #ifndef PRODUCT
3089     if (_vector_loop_debug) {
3090       tty->print_cr("SuperWord::find_phi_for_mem_dep _clone_map.gen(mem->_idx)=%d",
3091                     _clone_map.gen(mem->_idx));
3092     }
3093 #endif
3094     return NULL; // does not depend on loop volatile node or depends on the same generation
3095   }
3096 
3097   //otherwise first node should depend on mem-phi
3098   Node* first = first_node(ld);
3099   assert(first->is_Load(), "must be Load");
3100   Node* phi = first->as_Load()->in(MemNode::Memory);
3101   if (!phi->is_Phi() || phi->bottom_type() != Type::MEMORY) {
3102 #ifndef PRODUCT
3103     if (_vector_loop_debug) {
3104       tty->print_cr("SuperWord::find_phi_for_mem_dep load is not vectorizable node, since it's `first` does not take input from mem phi");
3105       ld->dump();
3106       first->dump();
3107     }
3108 #endif
3109     return NULL;
3110   }
3111 
3112   Node* tail = 0;
3113   for (int m = 0; m < _mem_slice_head.length(); m++) {
3114     if (_mem_slice_head.at(m) == phi) {
3115       tail = _mem_slice_tail.at(m);
3116     }
3117   }
3118   if (tail == 0) { //test that found phi is in the list  _mem_slice_head
3119 #ifndef PRODUCT
3120     if (_vector_loop_debug) {
3121       tty->print_cr("SuperWord::find_phi_for_mem_dep load %d is not vectorizable node, its phi %d is not _mem_slice_head",
3122                     ld->_idx, phi->_idx);
3123       ld->dump();
3124       phi->dump();
3125     }
3126 #endif
3127     return NULL;
3128   }
3129 
3130   // now all conditions are met
3131   return phi;
3132 }
3133 
3134 Node* SuperWord::first_node(Node* nd) {
3135   for (int ii = 0; ii < _iteration_first.length(); ii++) {
3136     Node* nnn = _iteration_first.at(ii);
3137     if (_clone_map.idx(nnn->_idx) == _clone_map.idx(nd->_idx)) {
3138 #ifndef PRODUCT
3139       if (_vector_loop_debug) {
3140         tty->print_cr("SuperWord::first_node: %d is the first iteration node for %d (_clone_map.idx(nnn->_idx) = %d)",
3141                       nnn->_idx, nd->_idx, _clone_map.idx(nnn->_idx));
3142       }
3143 #endif
3144       return nnn;
3145     }
3146   }
3147 
3148 #ifndef PRODUCT
3149   if (_vector_loop_debug) {
3150     tty->print_cr("SuperWord::first_node: did not find first iteration node for %d (_clone_map.idx(nd->_idx)=%d)",
3151                   nd->_idx, _clone_map.idx(nd->_idx));
3152   }
3153 #endif
3154   return 0;
3155 }
3156 
3157 Node* SuperWord::last_node(Node* nd) {
3158   for (int ii = 0; ii < _iteration_last.length(); ii++) {
3159     Node* nnn = _iteration_last.at(ii);
3160     if (_clone_map.idx(nnn->_idx) == _clone_map.idx(nd->_idx)) {
3161 #ifndef PRODUCT
3162       if (_vector_loop_debug) {
3163         tty->print_cr("SuperWord::last_node _clone_map.idx(nnn->_idx)=%d, _clone_map.idx(nd->_idx)=%d",
3164                       _clone_map.idx(nnn->_idx), _clone_map.idx(nd->_idx));
3165       }
3166 #endif
3167       return nnn;
3168     }
3169   }
3170   return 0;
3171 }
3172 
3173 int SuperWord::mark_generations() {
3174   Node *ii_err = 0, *tail_err;
3175   for (int i = 0; i < _mem_slice_head.length(); i++) {
3176     Node* phi  = _mem_slice_head.at(i);
3177     assert(phi->is_Phi(), "must be phi");
3178 
3179     Node* tail = _mem_slice_tail.at(i);
3180     if (_ii_last == -1) {
3181       tail_err = tail;
3182       _ii_last = _clone_map.gen(tail->_idx);
3183     }
3184     else if (_ii_last != _clone_map.gen(tail->_idx)) {
3185 #ifndef PRODUCT
3186       if (TraceSuperWord && Verbose) {
3187         tty->print_cr("SuperWord::mark_generations _ii_last error - found different generations in two tail nodes ");
3188         tail->dump();
3189         tail_err->dump();
3190       }
3191 #endif
3192       return -1;
3193     }
3194 
3195     // find first iteration in the loop
3196     for (DUIterator_Fast imax, i = phi->fast_outs(imax); i < imax; i++) {
3197       Node* ii = phi->fast_out(i);
3198       if (in_bb(ii) && ii->is_Store()) { // we speculate that normally Stores of one and one only generation have deps from mem phi
3199         if (_ii_first == -1) {
3200           ii_err = ii;
3201           _ii_first = _clone_map.gen(ii->_idx);
3202         } else if (_ii_first != _clone_map.gen(ii->_idx)) {
3203 #ifndef PRODUCT
3204           if (TraceSuperWord && Verbose) {
3205             tty->print_cr("SuperWord::mark_generations _ii_first error - found different generations in two nodes ");
3206             ii->dump();
3207             ii_err->dump();
3208           }
3209 #endif
3210           return -1; // this phi has Stores from different generations of unroll and cannot be simd/vectorized
3211         }
3212       }
3213     }//for (DUIterator_Fast imax,
3214   }//for (int i...
3215 
3216   if (_ii_first == -1 || _ii_last == -1) {
3217 #ifndef PRODUCT
3218     if (TraceSuperWord && Verbose) {
3219       tty->print_cr("SuperWord::mark_generations unknown error, something vent wrong");
3220     }
3221 #endif
3222     return -1; // something vent wrong
3223   }
3224   // collect nodes in the first and last generations
3225   assert(_iteration_first.length() == 0, "_iteration_first must be empty");
3226   assert(_iteration_last.length() == 0, "_iteration_last must be empty");
3227   for (int j = 0; j < _block.length(); j++) {
3228     Node* n = _block.at(j);
3229     node_idx_t gen = _clone_map.gen(n->_idx);
3230     if ((signed)gen == _ii_first) {
3231       _iteration_first.push(n);
3232     } else if ((signed)gen == _ii_last) {
3233       _iteration_last.push(n);
3234     }
3235   }
3236 
3237   // building order of iterations
3238   assert(_ii_order.length() == 0, "should be empty");
3239   if (ii_err != 0) {
3240     assert(in_bb(ii_err) && ii_err->is_Store(), "should be Store in bb");
3241     Node* nd = ii_err;
3242     while(_clone_map.gen(nd->_idx) != _ii_last) {
3243       _ii_order.push(_clone_map.gen(nd->_idx));
3244       bool found = false;
3245       for (DUIterator_Fast imax, i = nd->fast_outs(imax); i < imax; i++) {
3246         Node* use = nd->fast_out(i);
3247         if (_clone_map.idx(use->_idx) == _clone_map.idx(nd->_idx) && use->as_Store()->in(MemNode::Memory) == nd) {
3248           found = true;
3249           nd = use;
3250           break;
3251         }
3252       }//for
3253 
3254       if (found == false) {
3255 #ifndef PRODUCT
3256         if (TraceSuperWord && Verbose) {
3257           tty->print_cr("SuperWord::mark_generations: Cannot build order of iterations - no dependent Store for %d", nd->_idx);
3258         }
3259 #endif
3260         _ii_order.clear();
3261         return -1;
3262       }
3263     } //while
3264     _ii_order.push(_clone_map.gen(nd->_idx));
3265   }
3266 
3267 #ifndef PRODUCT
3268   if (_vector_loop_debug) {
3269     tty->print_cr("SuperWord::mark_generations");
3270     tty->print_cr("First generation (%d) nodes:", _ii_first);
3271     for (int ii = 0; ii < _iteration_first.length(); ii++)  _iteration_first.at(ii)->dump();
3272     tty->print_cr("Last generation (%d) nodes:", _ii_last);
3273     for (int ii = 0; ii < _iteration_last.length(); ii++)  _iteration_last.at(ii)->dump();
3274     tty->print_cr(" ");
3275 
3276     tty->print("SuperWord::List of generations: ");
3277     for (int jj = 0; jj < _ii_order.length(); ++jj) {
3278       tty->print("%d:%d ", jj, _ii_order.at(jj));
3279     }
3280     tty->print_cr(" ");
3281   }
3282 #endif
3283 
3284   return _ii_first;
3285 }
3286 
3287 bool SuperWord::fix_commutative_inputs(Node* gold, Node* fix) {
3288   assert(gold->is_Add() && fix->is_Add() || gold->is_Mul() && fix->is_Mul(), "should be only Add or Mul nodes");
3289   assert(_clone_map.idx(gold->_idx) == _clone_map.idx(fix->_idx), "should be clones of the same node");
3290   Node* gin1 = gold->in(1);
3291   Node* gin2 = gold->in(2);
3292   Node* fin1 = fix->in(1);
3293   Node* fin2 = fix->in(2);
3294   bool swapped = false;
3295 
3296   if (in_bb(gin1) && in_bb(gin2) && in_bb(fin1) && in_bb(fin1)) {
3297     if (_clone_map.idx(gin1->_idx) == _clone_map.idx(fin1->_idx) &&
3298         _clone_map.idx(gin2->_idx) == _clone_map.idx(fin2->_idx)) {
3299       return true; // nothing to fix
3300     }
3301     if (_clone_map.idx(gin1->_idx) == _clone_map.idx(fin2->_idx) &&
3302         _clone_map.idx(gin2->_idx) == _clone_map.idx(fin1->_idx)) {
3303       fix->swap_edges(1, 2);
3304       swapped = true;
3305     }
3306   }
3307   // at least one input comes from outside of bb
3308   if (gin1->_idx == fin1->_idx)  {
3309     return true; // nothing to fix
3310   }
3311   if (!swapped && (gin1->_idx == fin2->_idx || gin2->_idx == fin1->_idx))  { //swapping is expensive, check condition first
3312     fix->swap_edges(1, 2);
3313     swapped = true;
3314   }
3315 
3316   if (swapped) {
3317 #ifndef PRODUCT
3318     if (_vector_loop_debug) {
3319       tty->print_cr("SuperWord::fix_commutative_inputs: fixed node %d", fix->_idx);
3320     }
3321 #endif
3322     return true;
3323   }
3324 
3325 #ifndef PRODUCT
3326   if (TraceSuperWord && Verbose) {
3327     tty->print_cr("SuperWord::fix_commutative_inputs: cannot fix node %d", fix->_idx);
3328   }
3329 #endif
3330   return false;
3331 }
3332 
3333 bool SuperWord::pack_parallel() {
3334 #ifndef PRODUCT
3335   if (_vector_loop_debug) {
3336     tty->print_cr("SuperWord::pack_parallel: START");
3337   }
3338 #endif
3339 
3340   _packset.clear();
3341 
3342   for (int ii = 0; ii < _iteration_first.length(); ii++) {
3343     Node* nd = _iteration_first.at(ii);
3344     if (in_bb(nd) && (nd->is_Load() || nd->is_Store() || nd->is_Add() || nd->is_Mul())) {
3345       Node_List* pk = new Node_List();
3346       pk->push(nd);
3347       for (int gen = 1; gen < _ii_order.length(); ++gen) {
3348         for (int kk = 0; kk < _block.length(); kk++) {
3349           Node* clone = _block.at(kk);
3350           if (_clone_map.idx(clone->_idx) == _clone_map.idx(nd->_idx) &&
3351               _clone_map.gen(clone->_idx) == _ii_order.at(gen)) {
3352             if (nd->is_Add() || nd->is_Mul()) {
3353               fix_commutative_inputs(nd, clone);
3354             }
3355             pk->push(clone);
3356             if (pk->size() == 4) {
3357               _packset.append(pk);
3358 #ifndef PRODUCT
3359               if (_vector_loop_debug) {
3360                 tty->print_cr("SuperWord::pack_parallel: added pack ");
3361                 pk->dump();
3362               }
3363 #endif
3364               if (_clone_map.gen(clone->_idx) != _ii_last) {
3365                 pk = new Node_List();
3366               }
3367             }
3368             break;
3369           }
3370         }
3371       }//for
3372     }//if
3373   }//for
3374 
3375 #ifndef PRODUCT
3376   if (_vector_loop_debug) {
3377     tty->print_cr("SuperWord::pack_parallel: END");
3378   }
3379 #endif
3380 
3381   return true;
3382 }
3383 
3384 bool SuperWord::hoist_loads_in_graph() {
3385   GrowableArray<Node*> loads;
3386 
3387 #ifndef PRODUCT
3388   if (_vector_loop_debug) {
3389     tty->print_cr("SuperWord::hoist_loads_in_graph: total number _mem_slice_head.length() = %d", _mem_slice_head.length());
3390   }
3391 #endif
3392 
3393   for (int i = 0; i < _mem_slice_head.length(); i++) {
3394     Node* n = _mem_slice_head.at(i);
3395     if ( !in_bb(n) || !n->is_Phi() || n->bottom_type() != Type::MEMORY) {
3396 #ifndef PRODUCT
3397       if (TraceSuperWord && Verbose) {
3398         tty->print_cr("SuperWord::hoist_loads_in_graph: skipping unexpected node n=%d", n->_idx);
3399       }
3400 #endif
3401       continue;
3402     }
3403 
3404 #ifndef PRODUCT
3405     if (_vector_loop_debug) {
3406       tty->print_cr("SuperWord::hoist_loads_in_graph: processing phi %d  = _mem_slice_head.at(%d);", n->_idx, i);
3407     }
3408 #endif
3409 
3410     for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3411       Node* ld = n->fast_out(i);
3412       if (ld->is_Load() && ld->as_Load()->in(MemNode::Memory) == n && in_bb(ld)) {
3413         for (int i = 0; i < _block.length(); i++) {
3414           Node* ld2 = _block.at(i);
3415           if (ld2->is_Load() &&
3416               _clone_map.idx(ld->_idx) == _clone_map.idx(ld2->_idx) &&
3417               _clone_map.gen(ld->_idx) != _clone_map.gen(ld2->_idx)) { // <= do not collect the first generation ld
3418 #ifndef PRODUCT
3419             if (_vector_loop_debug) {
3420               tty->print_cr("SuperWord::hoist_loads_in_graph: will try to hoist load ld2->_idx=%d, cloned from %d (ld->_idx=%d)",
3421                             ld2->_idx, _clone_map.idx(ld->_idx), ld->_idx);
3422             }
3423 #endif
3424             // could not do on-the-fly, since iterator is immutable
3425             loads.push(ld2);
3426           }
3427         }// for
3428       }//if
3429     }//for (DUIterator_Fast imax,
3430   }//for (int i = 0; i
3431 
3432   for (int i = 0; i < loads.length(); i++) {
3433     LoadNode* ld = loads.at(i)->as_Load();
3434     Node* phi = find_phi_for_mem_dep(ld);
3435     if (phi != NULL) {
3436 #ifndef PRODUCT
3437       if (_vector_loop_debug) {
3438         tty->print_cr("SuperWord::hoist_loads_in_graph replacing MemNode::Memory(%d) edge in %d with one from %d",
3439                       MemNode::Memory, ld->_idx, phi->_idx);
3440       }
3441 #endif
3442       _igvn.replace_input_of(ld, MemNode::Memory, phi);
3443     }
3444   }//for
3445 
3446   restart(); // invalidate all basic structures, since we rebuilt the graph
3447 
3448 #ifndef PRODUCT
3449   if (TraceSuperWord && Verbose) {
3450     tty->print_cr("\nSuperWord::hoist_loads_in_graph() the graph was rebuilt, all structures invalidated and need rebuild");
3451   }
3452 #endif
3453   return true;
3454 }
3455