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   // We only re-enter slp when we vector mapped a queried loop and we want to
 104   // continue unrolling, in this case, slp is not subsequently done.
 105   if (cl->ignore_slp()) return;
 106 
 107   // Check for pre-loop ending with CountedLoopEnd(Bool(Cmp(x,Opaque1(limit))))
 108   CountedLoopEndNode* pre_end = get_pre_loop_end(cl);
 109   if (pre_end == NULL) return;
 110   Node *pre_opaq1 = pre_end->limit();
 111   if (pre_opaq1->Opcode() != Op_Opaque1) return;
 112 
 113   init(); // initialize data structures
 114 
 115   set_lpt(lpt);
 116   set_lp(cl);
 117 
 118   // For now, define one block which is the entire loop body
 119   set_bb(cl);
 120 
 121   if (do_optimization) {
 122     assert(_packset.length() == 0, "packset must be empty");
 123     SLP_extract();
 124   }
 125 }
 126 
 127 //------------------------------early unrolling analysis------------------------------
 128 void SuperWord::unrolling_analysis(int &local_loop_unroll_factor) {
 129   bool is_slp = true;
 130   ResourceMark rm;
 131   size_t ignored_size = lpt()->_body.size();
 132   int *ignored_loop_nodes = NEW_RESOURCE_ARRAY(int, ignored_size);
 133   Node_Stack nstack((int)ignored_size);
 134   CountedLoopNode *cl = lpt()->_head->as_CountedLoop();
 135   Node *cl_exit = cl->loopexit();
 136 
 137   // First clear the entries
 138   for (uint i = 0; i < lpt()->_body.size(); i++) {
 139     ignored_loop_nodes[i] = -1;
 140   }
 141 
 142   int max_vector = Matcher::max_vector_size(T_INT);
 143 
 144   // Process the loop, some/all of the stack entries will not be in order, ergo
 145   // need to preprocess the ignored initial state before we process the loop
 146   for (uint i = 0; i < lpt()->_body.size(); i++) {
 147     Node* n = lpt()->_body.at(i);
 148     if (n == cl->incr() ||
 149       n->is_reduction() ||
 150       n->is_AddP() ||
 151       n->is_Cmp() ||
 152       n->is_IfTrue() ||
 153       n->is_CountedLoop() ||
 154       (n == cl_exit)) {
 155       ignored_loop_nodes[i] = n->_idx;
 156       continue;
 157     }
 158 
 159     if (n->is_If()) {
 160       IfNode *iff = n->as_If();
 161       if (iff->_fcnt != COUNT_UNKNOWN && iff->_prob != PROB_UNKNOWN) {
 162         if (lpt()->is_loop_exit(iff)) {
 163           ignored_loop_nodes[i] = n->_idx;
 164           continue;
 165         }
 166       }
 167     }
 168 
 169     if (n->is_Phi() && (n->bottom_type() == Type::MEMORY)) {
 170       Node* n_tail = n->in(LoopNode::LoopBackControl);
 171       if (n_tail != n->in(LoopNode::EntryControl)) {
 172         if (!n_tail->is_Mem()) {
 173           is_slp = false;
 174           break;
 175         }
 176       }
 177     }
 178 
 179     // This must happen after check of phi/if
 180     if (n->is_Phi() || n->is_If()) {
 181       ignored_loop_nodes[i] = n->_idx;
 182       continue;
 183     }
 184 
 185     if (n->is_LoadStore() || n->is_MergeMem() ||
 186       (n->is_Proj() && !n->as_Proj()->is_CFG())) {
 187       is_slp = false;
 188       break;
 189     }
 190 
 191     if (n->is_Mem()) {
 192       MemNode* current = n->as_Mem();
 193       BasicType bt = current->memory_type();
 194       if (is_java_primitive(bt) == false) {
 195         ignored_loop_nodes[i] = n->_idx;
 196         continue;
 197       }
 198       Node* adr = n->in(MemNode::Address);
 199       Node* n_ctrl = _phase->get_ctrl(adr);
 200 
 201       // save a queue of post process nodes
 202       if (n_ctrl != NULL && lpt()->is_member(_phase->get_loop(n_ctrl))) {
 203         // Process the memory expression
 204         int stack_idx = 0;
 205         bool have_side_effects = true;
 206         if (adr->is_AddP() == false) {
 207           nstack.push(adr, stack_idx++);
 208         } else {
 209           // Mark the components of the memory operation in nstack
 210           SWPointer p1(current, this, &nstack, true);
 211           have_side_effects = p1.node_stack()->is_nonempty();
 212         }
 213 
 214         // Process the pointer stack
 215         while (have_side_effects) {
 216           Node* pointer_node = nstack.node();
 217           for (uint j = 0; j < lpt()->_body.size(); j++) {
 218             Node* cur_node = lpt()->_body.at(j);
 219             if (cur_node == pointer_node) {
 220               ignored_loop_nodes[j] = cur_node->_idx;
 221               break;
 222             }
 223           }
 224           nstack.pop();
 225           have_side_effects = nstack.is_nonempty();
 226         }
 227       }
 228     }
 229   }
 230 
 231   if (is_slp) {
 232     // Now we try to find the maximum supported consistent vector which the machine
 233     // description can use
 234     for (uint i = 0; i < lpt()->_body.size(); i++) {
 235       if (ignored_loop_nodes[i] != -1) continue;
 236 
 237       BasicType bt;
 238       Node* n = lpt()->_body.at(i);
 239       if (n->is_Store()) {
 240         bt = n->as_Mem()->memory_type();
 241       } else {
 242         bt = n->bottom_type()->basic_type();
 243       }
 244 
 245       int cur_max_vector = Matcher::max_vector_size(bt);
 246 
 247       // If a max vector exists which is not larger than _local_loop_unroll_factor
 248       // stop looking, we already have the max vector to map to.
 249       if (cur_max_vector < local_loop_unroll_factor) {
 250         is_slp = false;
 251         NOT_PRODUCT(if (TraceSuperWordLoopUnrollAnalysis) tty->print_cr("slp analysis fails: unroll limit greater than max vector\n"));
 252         break;
 253       }
 254 
 255       // Map the maximal common vector
 256       if (VectorNode::implemented(n->Opcode(), cur_max_vector, bt)) {
 257         if (cur_max_vector < max_vector) {
 258           max_vector = cur_max_vector;
 259         }
 260       }
 261     }
 262     if (is_slp) {
 263       local_loop_unroll_factor = max_vector;
 264       cl->mark_passed_slp();
 265     }
 266     cl->mark_was_slp();
 267     cl->set_slp_max_unroll(local_loop_unroll_factor);
 268   }
 269 }
 270 
 271 //------------------------------SLP_extract---------------------------
 272 // Extract the superword level parallelism
 273 //
 274 // 1) A reverse post-order of nodes in the block is constructed.  By scanning
 275 //    this list from first to last, all definitions are visited before their uses.
 276 //
 277 // 2) A point-to-point dependence graph is constructed between memory references.
 278 //    This simplies the upcoming "independence" checker.
 279 //
 280 // 3) The maximum depth in the node graph from the beginning of the block
 281 //    to each node is computed.  This is used to prune the graph search
 282 //    in the independence checker.
 283 //
 284 // 4) For integer types, the necessary bit width is propagated backwards
 285 //    from stores to allow packed operations on byte, char, and short
 286 //    integers.  This reverses the promotion to type "int" that javac
 287 //    did for operations like: char c1,c2,c3;  c1 = c2 + c3.
 288 //
 289 // 5) One of the memory references is picked to be an aligned vector reference.
 290 //    The pre-loop trip count is adjusted to align this reference in the
 291 //    unrolled body.
 292 //
 293 // 6) The initial set of pack pairs is seeded with memory references.
 294 //
 295 // 7) The set of pack pairs is extended by following use->def and def->use links.
 296 //
 297 // 8) The pairs are combined into vector sized packs.
 298 //
 299 // 9) Reorder the memory slices to co-locate members of the memory packs.
 300 //
 301 // 10) Generate ideal vector nodes for the final set of packs and where necessary,
 302 //    inserting scalar promotion, vector creation from multiple scalars, and
 303 //    extraction of scalar values from vectors.
 304 //
 305 void SuperWord::SLP_extract() {
 306 
 307 #ifndef PRODUCT
 308   if (_do_vector_loop && TraceSuperWord) {
 309     tty->print("SuperWord::SLP_extract\n");
 310     tty->print("input loop\n");
 311     _lpt->dump_head();
 312     _lpt->dump();
 313     for (uint i = 0; i < _lpt->_body.size(); i++) {
 314       _lpt->_body.at(i)->dump();
 315     }
 316   }
 317 #endif
 318   // Ready the block
 319   if (!construct_bb()) {
 320     return; // Exit if no interesting nodes or complex graph.
 321   }
 322   // build    _dg, _disjoint_ptrs
 323   dependence_graph();
 324 
 325   // compute function depth(Node*)
 326   compute_max_depth();
 327 
 328   if (_do_vector_loop) {
 329     if (mark_generations() != -1) {
 330       hoist_loads_in_graph(); // this only rebuild the graph; all basic structs need rebuild explicitly
 331 
 332       if (!construct_bb()) {
 333         return; // Exit if no interesting nodes or complex graph.
 334       }
 335       dependence_graph();
 336       compute_max_depth();
 337     }
 338 
 339 #ifndef PRODUCT
 340     if (TraceSuperWord) {
 341       tty->print_cr("\nSuperWord::_do_vector_loop: graph after hoist_loads_in_graph");
 342       _lpt->dump_head();
 343       for (int j = 0; j < _block.length(); j++) {
 344         Node* n = _block.at(j);
 345         int d = depth(n);
 346         for (int i = 0;  i < d; i++) tty->print("%s", "  ");
 347         tty->print("%d :", d);
 348         n->dump();
 349       }
 350     }
 351 #endif
 352   }
 353 
 354   compute_vector_element_type();
 355 
 356   // Attempt vectorization
 357 
 358   find_adjacent_refs();
 359 
 360   extend_packlist();
 361 
 362   if (_do_vector_loop) {
 363     if (_packset.length() == 0) {
 364 #ifndef PRODUCT
 365       if (TraceSuperWord) {
 366         tty->print_cr("\nSuperWord::_do_vector_loop DFA could not build packset, now trying to build anyway");
 367       }
 368 #endif
 369       pack_parallel();
 370     }
 371   }
 372 
 373   combine_packs();
 374 
 375   construct_my_pack_map();
 376 
 377   filter_packs();
 378 
 379   schedule();
 380 
 381   output();
 382 }
 383 
 384 //------------------------------find_adjacent_refs---------------------------
 385 // Find the adjacent memory references and create pack pairs for them.
 386 // This is the initial set of packs that will then be extended by
 387 // following use->def and def->use links.  The align positions are
 388 // assigned relative to the reference "align_to_ref"
 389 void SuperWord::find_adjacent_refs() {
 390   // Get list of memory operations
 391   Node_List memops;
 392   for (int i = 0; i < _block.length(); i++) {
 393     Node* n = _block.at(i);
 394     if (n->is_Mem() && !n->is_LoadStore() && in_bb(n) &&
 395         is_java_primitive(n->as_Mem()->memory_type())) {
 396       int align = memory_alignment(n->as_Mem(), 0);
 397       if (align != bottom_align) {
 398         memops.push(n);
 399       }
 400     }
 401   }
 402 
 403   Node_List align_to_refs;
 404   int best_iv_adjustment = 0;
 405   MemNode* best_align_to_mem_ref = NULL;
 406 
 407   while (memops.size() != 0) {
 408     // Find a memory reference to align to.
 409     MemNode* mem_ref = find_align_to_ref(memops);
 410     if (mem_ref == NULL) break;
 411     align_to_refs.push(mem_ref);
 412     int iv_adjustment = get_iv_adjustment(mem_ref);
 413 
 414     if (best_align_to_mem_ref == NULL) {
 415       // Set memory reference which is the best from all memory operations
 416       // to be used for alignment. The pre-loop trip count is modified to align
 417       // this reference to a vector-aligned address.
 418       best_align_to_mem_ref = mem_ref;
 419       best_iv_adjustment = iv_adjustment;
 420     }
 421 
 422     SWPointer align_to_ref_p(mem_ref, this, NULL, false);
 423     // Set alignment relative to "align_to_ref" for all related memory operations.
 424     for (int i = memops.size() - 1; i >= 0; i--) {
 425       MemNode* s = memops.at(i)->as_Mem();
 426       if (isomorphic(s, mem_ref)) {
 427         SWPointer p2(s, this, NULL, false);
 428         if (p2.comparable(align_to_ref_p)) {
 429           int align = memory_alignment(s, iv_adjustment);
 430           set_alignment(s, align);
 431         }
 432       }
 433     }
 434 
 435     // Create initial pack pairs of memory operations for which
 436     // alignment is set and vectors will be aligned.
 437     bool create_pack = true;
 438     if (memory_alignment(mem_ref, best_iv_adjustment) == 0 || _do_vector_loop) {
 439       if (!Matcher::misaligned_vectors_ok()) {
 440         int vw = vector_width(mem_ref);
 441         int vw_best = vector_width(best_align_to_mem_ref);
 442         if (vw > vw_best) {
 443           // Do not vectorize a memory access with more elements per vector
 444           // if unaligned memory access is not allowed because number of
 445           // iterations in pre-loop will be not enough to align it.
 446           create_pack = false;
 447         } else {
 448           SWPointer p2(best_align_to_mem_ref, this, NULL, false);
 449           if (align_to_ref_p.invar() != p2.invar()) {
 450             // Do not vectorize memory accesses with different invariants
 451             // if unaligned memory accesses are not allowed.
 452             create_pack = false;
 453           }
 454         }
 455       }
 456     } else {
 457       if (same_velt_type(mem_ref, best_align_to_mem_ref)) {
 458         // Can't allow vectorization of unaligned memory accesses with the
 459         // same type since it could be overlapped accesses to the same array.
 460         create_pack = false;
 461       } else {
 462         // Allow independent (different type) unaligned memory operations
 463         // if HW supports them.
 464         if (!Matcher::misaligned_vectors_ok()) {
 465           create_pack = false;
 466         } else {
 467           // Check if packs of the same memory type but
 468           // with a different alignment were created before.
 469           for (uint i = 0; i < align_to_refs.size(); i++) {
 470             MemNode* mr = align_to_refs.at(i)->as_Mem();
 471             if (same_velt_type(mr, mem_ref) &&
 472                 memory_alignment(mr, iv_adjustment) != 0)
 473               create_pack = false;
 474           }
 475         }
 476       }
 477     }
 478     if (create_pack) {
 479       for (uint i = 0; i < memops.size(); i++) {
 480         Node* s1 = memops.at(i);
 481         int align = alignment(s1);
 482         if (align == top_align) continue;
 483         for (uint j = 0; j < memops.size(); j++) {
 484           Node* s2 = memops.at(j);
 485           if (alignment(s2) == top_align) continue;
 486           if (s1 != s2 && are_adjacent_refs(s1, s2)) {
 487             if (stmts_can_pack(s1, s2, align)) {
 488               Node_List* pair = new Node_List();
 489               pair->push(s1);
 490               pair->push(s2);
 491               if (!_do_vector_loop || _clone_map.idx(s1->_idx) == _clone_map.idx(s2->_idx)) {
 492                 _packset.append(pair);
 493               }
 494             }
 495           }
 496         }
 497       }
 498     } else { // Don't create unaligned pack
 499       // First, remove remaining memory ops of the same type from the list.
 500       for (int i = memops.size() - 1; i >= 0; i--) {
 501         MemNode* s = memops.at(i)->as_Mem();
 502         if (same_velt_type(s, mem_ref)) {
 503           memops.remove(i);
 504         }
 505       }
 506 
 507       // Second, remove already constructed packs of the same type.
 508       for (int i = _packset.length() - 1; i >= 0; i--) {
 509         Node_List* p = _packset.at(i);
 510         MemNode* s = p->at(0)->as_Mem();
 511         if (same_velt_type(s, mem_ref)) {
 512           remove_pack_at(i);
 513         }
 514       }
 515 
 516       // If needed find the best memory reference for loop alignment again.
 517       if (same_velt_type(mem_ref, best_align_to_mem_ref)) {
 518         // Put memory ops from remaining packs back on memops list for
 519         // the best alignment search.
 520         uint orig_msize = memops.size();
 521         for (int i = 0; i < _packset.length(); i++) {
 522           Node_List* p = _packset.at(i);
 523           MemNode* s = p->at(0)->as_Mem();
 524           assert(!same_velt_type(s, mem_ref), "sanity");
 525           memops.push(s);
 526         }
 527         MemNode* best_align_to_mem_ref = find_align_to_ref(memops);
 528         if (best_align_to_mem_ref == NULL) break;
 529         best_iv_adjustment = get_iv_adjustment(best_align_to_mem_ref);
 530         // Restore list.
 531         while (memops.size() > orig_msize)
 532           (void)memops.pop();
 533       }
 534     } // unaligned memory accesses
 535 
 536     // Remove used mem nodes.
 537     for (int i = memops.size() - 1; i >= 0; i--) {
 538       MemNode* m = memops.at(i)->as_Mem();
 539       if (alignment(m) != top_align) {
 540         memops.remove(i);
 541       }
 542     }
 543 
 544   } // while (memops.size() != 0
 545   set_align_to_ref(best_align_to_mem_ref);
 546 
 547 #ifndef PRODUCT
 548   if (TraceSuperWord) {
 549     tty->print_cr("\nAfter find_adjacent_refs");
 550     print_packset();
 551   }
 552 #endif
 553 }
 554 
 555 //------------------------------find_align_to_ref---------------------------
 556 // Find a memory reference to align the loop induction variable to.
 557 // Looks first at stores then at loads, looking for a memory reference
 558 // with the largest number of references similar to it.
 559 MemNode* SuperWord::find_align_to_ref(Node_List &memops) {
 560   GrowableArray<int> cmp_ct(arena(), memops.size(), memops.size(), 0);
 561 
 562   // Count number of comparable memory ops
 563   for (uint i = 0; i < memops.size(); i++) {
 564     MemNode* s1 = memops.at(i)->as_Mem();
 565     SWPointer p1(s1, this, NULL, false);
 566     // Discard if pre loop can't align this reference
 567     if (!ref_is_alignable(p1)) {
 568       *cmp_ct.adr_at(i) = 0;
 569       continue;
 570     }
 571     for (uint j = i+1; j < memops.size(); j++) {
 572       MemNode* s2 = memops.at(j)->as_Mem();
 573       if (isomorphic(s1, s2)) {
 574         SWPointer p2(s2, this, NULL, false);
 575         if (p1.comparable(p2)) {
 576           (*cmp_ct.adr_at(i))++;
 577           (*cmp_ct.adr_at(j))++;
 578         }
 579       }
 580     }
 581   }
 582 
 583   // Find Store (or Load) with the greatest number of "comparable" references,
 584   // biggest vector size, smallest data size and smallest iv offset.
 585   int max_ct        = 0;
 586   int max_vw        = 0;
 587   int max_idx       = -1;
 588   int min_size      = max_jint;
 589   int min_iv_offset = max_jint;
 590   for (uint j = 0; j < memops.size(); j++) {
 591     MemNode* s = memops.at(j)->as_Mem();
 592     if (s->is_Store()) {
 593       int vw = vector_width_in_bytes(s);
 594       assert(vw > 1, "sanity");
 595       SWPointer p(s, this, NULL, false);
 596       if (cmp_ct.at(j) >  max_ct ||
 597           cmp_ct.at(j) == max_ct &&
 598             (vw >  max_vw ||
 599              vw == max_vw &&
 600               (data_size(s) <  min_size ||
 601                data_size(s) == min_size &&
 602                  (p.offset_in_bytes() < min_iv_offset)))) {
 603         max_ct = cmp_ct.at(j);
 604         max_vw = vw;
 605         max_idx = j;
 606         min_size = data_size(s);
 607         min_iv_offset = p.offset_in_bytes();
 608       }
 609     }
 610   }
 611   // If no stores, look at loads
 612   if (max_ct == 0) {
 613     for (uint j = 0; j < memops.size(); j++) {
 614       MemNode* s = memops.at(j)->as_Mem();
 615       if (s->is_Load()) {
 616         int vw = vector_width_in_bytes(s);
 617         assert(vw > 1, "sanity");
 618         SWPointer p(s, this, NULL, false);
 619         if (cmp_ct.at(j) >  max_ct ||
 620             cmp_ct.at(j) == max_ct &&
 621               (vw >  max_vw ||
 622                vw == max_vw &&
 623                 (data_size(s) <  min_size ||
 624                  data_size(s) == min_size &&
 625                    (p.offset_in_bytes() < min_iv_offset)))) {
 626           max_ct = cmp_ct.at(j);
 627           max_vw = vw;
 628           max_idx = j;
 629           min_size = data_size(s);
 630           min_iv_offset = p.offset_in_bytes();
 631         }
 632       }
 633     }
 634   }
 635 
 636 #ifdef ASSERT
 637   if (TraceSuperWord && Verbose) {
 638     tty->print_cr("\nVector memops after find_align_to_ref");
 639     for (uint i = 0; i < memops.size(); i++) {
 640       MemNode* s = memops.at(i)->as_Mem();
 641       s->dump();
 642     }
 643   }
 644 #endif
 645 
 646   if (max_ct > 0) {
 647 #ifdef ASSERT
 648     if (TraceSuperWord) {
 649       tty->print("\nVector align to node: ");
 650       memops.at(max_idx)->as_Mem()->dump();
 651     }
 652 #endif
 653     return memops.at(max_idx)->as_Mem();
 654   }
 655   return NULL;
 656 }
 657 
 658 //------------------------------ref_is_alignable---------------------------
 659 // Can the preloop align the reference to position zero in the vector?
 660 bool SuperWord::ref_is_alignable(SWPointer& p) {
 661   if (!p.has_iv()) {
 662     return true;   // no induction variable
 663   }
 664   CountedLoopEndNode* pre_end = get_pre_loop_end(lp()->as_CountedLoop());
 665   assert(pre_end != NULL, "we must have a correct pre-loop");
 666   assert(pre_end->stride_is_con(), "pre loop stride is constant");
 667   int preloop_stride = pre_end->stride_con();
 668 
 669   int span = preloop_stride * p.scale_in_bytes();
 670   int mem_size = p.memory_size();
 671   int offset   = p.offset_in_bytes();
 672   // Stride one accesses are alignable if offset is aligned to memory operation size.
 673   // Offset can be unaligned when UseUnalignedAccesses is used.
 674   if (ABS(span) == mem_size && (ABS(offset) % mem_size) == 0) {
 675     return true;
 676   }
 677   // If the initial offset from start of the object is computable,
 678   // check if the pre-loop can align the final offset accordingly.
 679   //
 680   // In other words: Can we find an i such that the offset
 681   // after i pre-loop iterations is aligned to vw?
 682   //   (init_offset + pre_loop) % vw == 0              (1)
 683   // where
 684   //   pre_loop = i * span
 685   // is the number of bytes added to the offset by i pre-loop iterations.
 686   //
 687   // For this to hold we need pre_loop to increase init_offset by
 688   //   pre_loop = vw - (init_offset % vw)
 689   //
 690   // This is only possible if pre_loop is divisible by span because each
 691   // pre-loop iteration increases the initial offset by 'span' bytes:
 692   //   (vw - (init_offset % vw)) % span == 0
 693   //
 694   int vw = vector_width_in_bytes(p.mem());
 695   assert(vw > 1, "sanity");
 696   Node* init_nd = pre_end->init_trip();
 697   if (init_nd->is_Con() && p.invar() == NULL) {
 698     int init = init_nd->bottom_type()->is_int()->get_con();
 699     int init_offset = init * p.scale_in_bytes() + offset;
 700     assert(init_offset >= 0, "positive offset from object start");
 701     if (vw % span == 0) {
 702       // If vm is a multiple of span, we use formula (1).
 703       if (span > 0) {
 704         return (vw - (init_offset % vw)) % span == 0;
 705       } else {
 706         assert(span < 0, "nonzero stride * scale");
 707         return (init_offset % vw) % -span == 0;
 708       }
 709     } else if (span % vw == 0) {
 710       // If span is a multiple of vw, we can simplify formula (1) to:
 711       //   (init_offset + i * span) % vw == 0
 712       //     =>
 713       //   (init_offset % vw) + ((i * span) % vw) == 0
 714       //     =>
 715       //   init_offset % vw == 0
 716       //
 717       // Because we add a multiple of vw to the initial offset, the final
 718       // offset is a multiple of vw if and only if init_offset is a multiple.
 719       //
 720       return (init_offset % vw) == 0;
 721     }
 722   }
 723   return false;
 724 }
 725 
 726 //---------------------------get_iv_adjustment---------------------------
 727 // Calculate loop's iv adjustment for this memory ops.
 728 int SuperWord::get_iv_adjustment(MemNode* mem_ref) {
 729   SWPointer align_to_ref_p(mem_ref, this, NULL, false);
 730   int offset = align_to_ref_p.offset_in_bytes();
 731   int scale  = align_to_ref_p.scale_in_bytes();
 732   int elt_size = align_to_ref_p.memory_size();
 733   int vw       = vector_width_in_bytes(mem_ref);
 734   assert(vw > 1, "sanity");
 735   int iv_adjustment;
 736   if (scale != 0) {
 737     int stride_sign = (scale * iv_stride()) > 0 ? 1 : -1;
 738     // At least one iteration is executed in pre-loop by default. As result
 739     // several iterations are needed to align memory operations in main-loop even
 740     // if offset is 0.
 741     int iv_adjustment_in_bytes = (stride_sign * vw - (offset % vw));
 742     assert(((ABS(iv_adjustment_in_bytes) % elt_size) == 0),
 743            err_msg_res("(%d) should be divisible by (%d)", iv_adjustment_in_bytes, elt_size));
 744     iv_adjustment = iv_adjustment_in_bytes/elt_size;
 745   } else {
 746     // This memory op is not dependent on iv (scale == 0)
 747     iv_adjustment = 0;
 748   }
 749 
 750 #ifndef PRODUCT
 751   if (TraceSuperWord)
 752     tty->print_cr("\noffset = %d iv_adjust = %d elt_size = %d scale = %d iv_stride = %d vect_size %d",
 753                   offset, iv_adjustment, elt_size, scale, iv_stride(), vw);
 754 #endif
 755   return iv_adjustment;
 756 }
 757 
 758 //---------------------------dependence_graph---------------------------
 759 // Construct dependency graph.
 760 // Add dependence edges to load/store nodes for memory dependence
 761 //    A.out()->DependNode.in(1) and DependNode.out()->B.prec(x)
 762 void SuperWord::dependence_graph() {
 763   // First, assign a dependence node to each memory node
 764   for (int i = 0; i < _block.length(); i++ ) {
 765     Node *n = _block.at(i);
 766     if (n->is_Mem() || n->is_Phi() && n->bottom_type() == Type::MEMORY) {
 767       _dg.make_node(n);
 768     }
 769   }
 770 
 771   // For each memory slice, create the dependences
 772   for (int i = 0; i < _mem_slice_head.length(); i++) {
 773     Node* n      = _mem_slice_head.at(i);
 774     Node* n_tail = _mem_slice_tail.at(i);
 775 
 776     // Get slice in predecessor order (last is first)
 777     mem_slice_preds(n_tail, n, _nlist);
 778 
 779 #ifndef PRODUCT
 780     if(TraceSuperWord && Verbose) {
 781       tty->print_cr("SuperWord::dependence_graph: built a new mem slice");
 782       for (int j = _nlist.length() - 1; j >= 0 ; j--) {
 783         _nlist.at(j)->dump();
 784       }
 785     }
 786 #endif
 787     // Make the slice dependent on the root
 788     DepMem* slice = _dg.dep(n);
 789     _dg.make_edge(_dg.root(), slice);
 790 
 791     // Create a sink for the slice
 792     DepMem* slice_sink = _dg.make_node(NULL);
 793     _dg.make_edge(slice_sink, _dg.tail());
 794 
 795     // Now visit each pair of memory ops, creating the edges
 796     for (int j = _nlist.length() - 1; j >= 0 ; j--) {
 797       Node* s1 = _nlist.at(j);
 798 
 799       // If no dependency yet, use slice
 800       if (_dg.dep(s1)->in_cnt() == 0) {
 801         _dg.make_edge(slice, s1);
 802       }
 803       SWPointer p1(s1->as_Mem(), this, NULL, false);
 804       bool sink_dependent = true;
 805       for (int k = j - 1; k >= 0; k--) {
 806         Node* s2 = _nlist.at(k);
 807         if (s1->is_Load() && s2->is_Load())
 808           continue;
 809         SWPointer p2(s2->as_Mem(), this, NULL, false);
 810 
 811         int cmp = p1.cmp(p2);
 812         if (SuperWordRTDepCheck &&
 813             p1.base() != p2.base() && p1.valid() && p2.valid()) {
 814           // Create a runtime check to disambiguate
 815           OrderedPair pp(p1.base(), p2.base());
 816           _disjoint_ptrs.append_if_missing(pp);
 817         } else if (!SWPointer::not_equal(cmp)) {
 818           // Possibly same address
 819           _dg.make_edge(s1, s2);
 820           sink_dependent = false;
 821         }
 822       }
 823       if (sink_dependent) {
 824         _dg.make_edge(s1, slice_sink);
 825       }
 826     }
 827 #ifndef PRODUCT
 828     if (TraceSuperWord) {
 829       tty->print_cr("\nDependence graph for slice: %d", n->_idx);
 830       for (int q = 0; q < _nlist.length(); q++) {
 831         _dg.print(_nlist.at(q));
 832       }
 833       tty->cr();
 834     }
 835 #endif
 836     _nlist.clear();
 837   }
 838 
 839 #ifndef PRODUCT
 840   if (TraceSuperWord) {
 841     tty->print_cr("\ndisjoint_ptrs: %s", _disjoint_ptrs.length() > 0 ? "" : "NONE");
 842     for (int r = 0; r < _disjoint_ptrs.length(); r++) {
 843       _disjoint_ptrs.at(r).print();
 844       tty->cr();
 845     }
 846     tty->cr();
 847   }
 848 #endif
 849 }
 850 
 851 //---------------------------mem_slice_preds---------------------------
 852 // Return a memory slice (node list) in predecessor order starting at "start"
 853 void SuperWord::mem_slice_preds(Node* start, Node* stop, GrowableArray<Node*> &preds) {
 854   assert(preds.length() == 0, "start empty");
 855   Node* n = start;
 856   Node* prev = NULL;
 857   while (true) {
 858     assert(in_bb(n), "must be in block");
 859     for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
 860       Node* out = n->fast_out(i);
 861       if (out->is_Load()) {
 862         if (in_bb(out)) {
 863           preds.push(out);
 864         }
 865       } else {
 866         // FIXME
 867         if (out->is_MergeMem() && !in_bb(out)) {
 868           // Either unrolling is causing a memory edge not to disappear,
 869           // or need to run igvn.optimize() again before SLP
 870         } else if (out->is_Phi() && out->bottom_type() == Type::MEMORY && !in_bb(out)) {
 871           // Ditto.  Not sure what else to check further.
 872         } else if (out->Opcode() == Op_StoreCM && out->in(MemNode::OopStore) == n) {
 873           // StoreCM has an input edge used as a precedence edge.
 874           // Maybe an issue when oop stores are vectorized.
 875         } else {
 876           assert(out == prev || prev == NULL, "no branches off of store slice");
 877         }
 878       }
 879     }
 880     if (n == stop) break;
 881     preds.push(n);
 882     prev = n;
 883     assert(n->is_Mem(), err_msg_res("unexpected node %s", n->Name()));
 884     n = n->in(MemNode::Memory);
 885   }
 886 }
 887 
 888 //------------------------------stmts_can_pack---------------------------
 889 // Can s1 and s2 be in a pack with s1 immediately preceding s2 and
 890 // s1 aligned at "align"
 891 bool SuperWord::stmts_can_pack(Node* s1, Node* s2, int align) {
 892 
 893   // Do not use superword for non-primitives
 894   BasicType bt1 = velt_basic_type(s1);
 895   BasicType bt2 = velt_basic_type(s2);
 896   if(!is_java_primitive(bt1) || !is_java_primitive(bt2))
 897     return false;
 898   if (Matcher::max_vector_size(bt1) < 2) {
 899     return false; // No vectors for this type
 900   }
 901 
 902   if (isomorphic(s1, s2)) {
 903     if (independent(s1, s2) || reduction(s1, s2)) {
 904       if (!exists_at(s1, 0) && !exists_at(s2, 1)) {
 905         if (!s1->is_Mem() || are_adjacent_refs(s1, s2)) {
 906           int s1_align = alignment(s1);
 907           int s2_align = alignment(s2);
 908           if (s1_align == top_align || s1_align == align) {
 909             if (s2_align == top_align || s2_align == align + data_size(s1)) {
 910               return true;
 911             }
 912           }
 913         }
 914       }
 915     }
 916   }
 917   return false;
 918 }
 919 
 920 //------------------------------exists_at---------------------------
 921 // Does s exist in a pack at position pos?
 922 bool SuperWord::exists_at(Node* s, uint pos) {
 923   for (int i = 0; i < _packset.length(); i++) {
 924     Node_List* p = _packset.at(i);
 925     if (p->at(pos) == s) {
 926       return true;
 927     }
 928   }
 929   return false;
 930 }
 931 
 932 //------------------------------are_adjacent_refs---------------------------
 933 // Is s1 immediately before s2 in memory?
 934 bool SuperWord::are_adjacent_refs(Node* s1, Node* s2) {
 935   if (!s1->is_Mem() || !s2->is_Mem()) return false;
 936   if (!in_bb(s1)    || !in_bb(s2))    return false;
 937 
 938   // Do not use superword for non-primitives
 939   if (!is_java_primitive(s1->as_Mem()->memory_type()) ||
 940       !is_java_primitive(s2->as_Mem()->memory_type())) {
 941     return false;
 942   }
 943 
 944   // FIXME - co_locate_pack fails on Stores in different mem-slices, so
 945   // only pack memops that are in the same alias set until that's fixed.
 946   if (_phase->C->get_alias_index(s1->as_Mem()->adr_type()) !=
 947       _phase->C->get_alias_index(s2->as_Mem()->adr_type()))
 948     return false;
 949   SWPointer p1(s1->as_Mem(), this, NULL, false);
 950   SWPointer p2(s2->as_Mem(), this, NULL, false);
 951   if (p1.base() != p2.base() || !p1.comparable(p2)) return false;
 952   int diff = p2.offset_in_bytes() - p1.offset_in_bytes();
 953   return diff == data_size(s1);
 954 }
 955 
 956 //------------------------------isomorphic---------------------------
 957 // Are s1 and s2 similar?
 958 bool SuperWord::isomorphic(Node* s1, Node* s2) {
 959   if (s1->Opcode() != s2->Opcode()) return false;
 960   if (s1->req() != s2->req()) return false;
 961   if (s1->in(0) != s2->in(0)) return false;
 962   if (!same_velt_type(s1, s2)) return false;
 963   return true;
 964 }
 965 
 966 //------------------------------independent---------------------------
 967 // Is there no data path from s1 to s2 or s2 to s1?
 968 bool SuperWord::independent(Node* s1, Node* s2) {
 969   //  assert(s1->Opcode() == s2->Opcode(), "check isomorphic first");
 970   int d1 = depth(s1);
 971   int d2 = depth(s2);
 972   if (d1 == d2) return s1 != s2;
 973   Node* deep    = d1 > d2 ? s1 : s2;
 974   Node* shallow = d1 > d2 ? s2 : s1;
 975 
 976   visited_clear();
 977 
 978   return independent_path(shallow, deep);
 979 }
 980 
 981 //------------------------------reduction---------------------------
 982 // Is there a data path between s1 and s2 and the nodes reductions?
 983 bool SuperWord::reduction(Node* s1, Node* s2) {
 984   bool retValue = false;
 985   int d1 = depth(s1);
 986   int d2 = depth(s2);
 987   if (d1 + 1 == d2) {
 988     if (s1->is_reduction() && s2->is_reduction()) {
 989       // This is an ordered set, so s1 should define s2
 990       for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
 991         Node* t1 = s1->fast_out(i);
 992         if (t1 == s2) {
 993           // both nodes are reductions and connected
 994           retValue = true;
 995         }
 996       }
 997     }
 998   }
 999 
1000   return retValue;
1001 }
1002 
1003 //------------------------------independent_path------------------------------
1004 // Helper for independent
1005 bool SuperWord::independent_path(Node* shallow, Node* deep, uint dp) {
1006   if (dp >= 1000) return false; // stop deep recursion
1007   visited_set(deep);
1008   int shal_depth = depth(shallow);
1009   assert(shal_depth <= depth(deep), "must be");
1010   for (DepPreds preds(deep, _dg); !preds.done(); preds.next()) {
1011     Node* pred = preds.current();
1012     if (in_bb(pred) && !visited_test(pred)) {
1013       if (shallow == pred) {
1014         return false;
1015       }
1016       if (shal_depth < depth(pred) && !independent_path(shallow, pred, dp+1)) {
1017         return false;
1018       }
1019     }
1020   }
1021   return true;
1022 }
1023 
1024 //------------------------------set_alignment---------------------------
1025 void SuperWord::set_alignment(Node* s1, Node* s2, int align) {
1026   set_alignment(s1, align);
1027   if (align == top_align || align == bottom_align) {
1028     set_alignment(s2, align);
1029   } else {
1030     set_alignment(s2, align + data_size(s1));
1031   }
1032 }
1033 
1034 //------------------------------data_size---------------------------
1035 int SuperWord::data_size(Node* s) {
1036   int bsize = type2aelembytes(velt_basic_type(s));
1037   assert(bsize != 0, "valid size");
1038   return bsize;
1039 }
1040 
1041 //------------------------------extend_packlist---------------------------
1042 // Extend packset by following use->def and def->use links from pack members.
1043 void SuperWord::extend_packlist() {
1044   bool changed;
1045   do {
1046     packset_sort(_packset.length());
1047     changed = false;
1048     for (int i = 0; i < _packset.length(); i++) {
1049       Node_List* p = _packset.at(i);
1050       changed |= follow_use_defs(p);
1051       changed |= follow_def_uses(p);
1052     }
1053   } while (changed);
1054 
1055   if (_race_possible) {
1056     for (int i = 0; i < _packset.length(); i++) {
1057       Node_List* p = _packset.at(i);
1058       order_def_uses(p);
1059     }
1060   }
1061 
1062 #ifndef PRODUCT
1063   if (TraceSuperWord) {
1064     tty->print_cr("\nAfter extend_packlist");
1065     print_packset();
1066   }
1067 #endif
1068 }
1069 
1070 //------------------------------follow_use_defs---------------------------
1071 // Extend the packset by visiting operand definitions of nodes in pack p
1072 bool SuperWord::follow_use_defs(Node_List* p) {
1073   assert(p->size() == 2, "just checking");
1074   Node* s1 = p->at(0);
1075   Node* s2 = p->at(1);
1076   assert(s1->req() == s2->req(), "just checking");
1077   assert(alignment(s1) + data_size(s1) == alignment(s2), "just checking");
1078 
1079   if (s1->is_Load()) return false;
1080 
1081   int align = alignment(s1);
1082   bool changed = false;
1083   int start = s1->is_Store() ? MemNode::ValueIn   : 1;
1084   int end   = s1->is_Store() ? MemNode::ValueIn+1 : s1->req();
1085   for (int j = start; j < end; j++) {
1086     Node* t1 = s1->in(j);
1087     Node* t2 = s2->in(j);
1088     if (!in_bb(t1) || !in_bb(t2))
1089       continue;
1090     if (stmts_can_pack(t1, t2, align)) {
1091       if (est_savings(t1, t2) >= 0) {
1092         Node_List* pair = new Node_List();
1093         pair->push(t1);
1094         pair->push(t2);
1095         _packset.append(pair);
1096         set_alignment(t1, t2, align);
1097         changed = true;
1098       }
1099     }
1100   }
1101   return changed;
1102 }
1103 
1104 //------------------------------follow_def_uses---------------------------
1105 // Extend the packset by visiting uses of nodes in pack p
1106 bool SuperWord::follow_def_uses(Node_List* p) {
1107   bool changed = false;
1108   Node* s1 = p->at(0);
1109   Node* s2 = p->at(1);
1110   assert(p->size() == 2, "just checking");
1111   assert(s1->req() == s2->req(), "just checking");
1112   assert(alignment(s1) + data_size(s1) == alignment(s2), "just checking");
1113 
1114   if (s1->is_Store()) return false;
1115 
1116   int align = alignment(s1);
1117   int savings = -1;
1118   int num_s1_uses = 0;
1119   Node* u1 = NULL;
1120   Node* u2 = NULL;
1121   for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
1122     Node* t1 = s1->fast_out(i);
1123     num_s1_uses++;
1124     if (!in_bb(t1)) continue;
1125     for (DUIterator_Fast jmax, j = s2->fast_outs(jmax); j < jmax; j++) {
1126       Node* t2 = s2->fast_out(j);
1127       if (!in_bb(t2)) continue;
1128       if (!opnd_positions_match(s1, t1, s2, t2))
1129         continue;
1130       if (stmts_can_pack(t1, t2, align)) {
1131         int my_savings = est_savings(t1, t2);
1132         if (my_savings > savings) {
1133           savings = my_savings;
1134           u1 = t1;
1135           u2 = t2;
1136         }
1137       }
1138     }
1139   }
1140   if (num_s1_uses > 1) {
1141     _race_possible = true;
1142   }
1143   if (savings >= 0) {
1144     Node_List* pair = new Node_List();
1145     pair->push(u1);
1146     pair->push(u2);
1147     _packset.append(pair);
1148     set_alignment(u1, u2, align);
1149     changed = true;
1150   }
1151   return changed;
1152 }
1153 
1154 //------------------------------order_def_uses---------------------------
1155 // For extended packsets, ordinally arrange uses packset by major component
1156 void SuperWord::order_def_uses(Node_List* p) {
1157   Node* s1 = p->at(0);
1158 
1159   if (s1->is_Store()) return;
1160 
1161   // reductions are always managed beforehand
1162   if (s1->is_reduction()) return;
1163 
1164   for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
1165     Node* t1 = s1->fast_out(i);
1166 
1167     // Only allow operand swap on commuting operations
1168     if (!t1->is_Add() && !t1->is_Mul()) {
1169       break;
1170     }
1171 
1172     // Now find t1's packset
1173     Node_List* p2 = NULL;
1174     for (int j = 0; j < _packset.length(); j++) {
1175       p2 = _packset.at(j);
1176       Node* first = p2->at(0);
1177       if (t1 == first) {
1178         break;
1179       }
1180       p2 = NULL;
1181     }
1182     // Arrange all sub components by the major component
1183     if (p2 != NULL) {
1184       for (uint j = 1; j < p->size(); j++) {
1185         Node* d1 = p->at(j);
1186         Node* u1 = p2->at(j);
1187         opnd_positions_match(s1, t1, d1, u1);
1188       }
1189     }
1190   }
1191 }
1192 
1193 //---------------------------opnd_positions_match-------------------------
1194 // Is the use of d1 in u1 at the same operand position as d2 in u2?
1195 bool SuperWord::opnd_positions_match(Node* d1, Node* u1, Node* d2, Node* u2) {
1196   // check reductions to see if they are marshalled to represent the reduction
1197   // operator in a specified opnd
1198   if (u1->is_reduction() && u2->is_reduction()) {
1199     // ensure reductions have phis and reduction definitions feeding the 1st operand
1200     Node* first = u1->in(2);
1201     if (first->is_Phi() || first->is_reduction()) {
1202       u1->swap_edges(1, 2);
1203     }
1204     // ensure reductions have phis and reduction definitions feeding the 1st operand
1205     first = u2->in(2);
1206     if (first->is_Phi() || first->is_reduction()) {
1207       u2->swap_edges(1, 2);
1208     }
1209     return true;
1210   }
1211 
1212   uint ct = u1->req();
1213   if (ct != u2->req()) return false;
1214   uint i1 = 0;
1215   uint i2 = 0;
1216   do {
1217     for (i1++; i1 < ct; i1++) if (u1->in(i1) == d1) break;
1218     for (i2++; i2 < ct; i2++) if (u2->in(i2) == d2) break;
1219     if (i1 != i2) {
1220       if ((i1 == (3-i2)) && (u2->is_Add() || u2->is_Mul())) {
1221         // Further analysis relies on operands position matching.
1222         u2->swap_edges(i1, i2);
1223       } else {
1224         return false;
1225       }
1226     }
1227   } while (i1 < ct);
1228   return true;
1229 }
1230 
1231 //------------------------------est_savings---------------------------
1232 // Estimate the savings from executing s1 and s2 as a pack
1233 int SuperWord::est_savings(Node* s1, Node* s2) {
1234   int save_in = 2 - 1; // 2 operations per instruction in packed form
1235 
1236   // inputs
1237   for (uint i = 1; i < s1->req(); i++) {
1238     Node* x1 = s1->in(i);
1239     Node* x2 = s2->in(i);
1240     if (x1 != x2) {
1241       if (are_adjacent_refs(x1, x2)) {
1242         save_in += adjacent_profit(x1, x2);
1243       } else if (!in_packset(x1, x2)) {
1244         save_in -= pack_cost(2);
1245       } else {
1246         save_in += unpack_cost(2);
1247       }
1248     }
1249   }
1250 
1251   // uses of result
1252   uint ct = 0;
1253   int save_use = 0;
1254   for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
1255     Node* s1_use = s1->fast_out(i);
1256     for (int j = 0; j < _packset.length(); j++) {
1257       Node_List* p = _packset.at(j);
1258       if (p->at(0) == s1_use) {
1259         for (DUIterator_Fast kmax, k = s2->fast_outs(kmax); k < kmax; k++) {
1260           Node* s2_use = s2->fast_out(k);
1261           if (p->at(p->size()-1) == s2_use) {
1262             ct++;
1263             if (are_adjacent_refs(s1_use, s2_use)) {
1264               save_use += adjacent_profit(s1_use, s2_use);
1265             }
1266           }
1267         }
1268       }
1269     }
1270   }
1271 
1272   if (ct < s1->outcnt()) save_use += unpack_cost(1);
1273   if (ct < s2->outcnt()) save_use += unpack_cost(1);
1274 
1275   return MAX2(save_in, save_use);
1276 }
1277 
1278 //------------------------------costs---------------------------
1279 int SuperWord::adjacent_profit(Node* s1, Node* s2) { return 2; }
1280 int SuperWord::pack_cost(int ct)   { return ct; }
1281 int SuperWord::unpack_cost(int ct) { return ct; }
1282 
1283 //------------------------------combine_packs---------------------------
1284 // Combine packs A and B with A.last == B.first into A.first..,A.last,B.second,..B.last
1285 void SuperWord::combine_packs() {
1286   bool changed = true;
1287   // Combine packs regardless max vector size.
1288   while (changed) {
1289     changed = false;
1290     for (int i = 0; i < _packset.length(); i++) {
1291       Node_List* p1 = _packset.at(i);
1292       if (p1 == NULL) continue;
1293       // Because of sorting we can start at i + 1
1294       for (int j = i + 1; j < _packset.length(); j++) {
1295         Node_List* p2 = _packset.at(j);
1296         if (p2 == NULL) continue;
1297         if (i == j) continue;
1298         if (p1->at(p1->size()-1) == p2->at(0)) {
1299           for (uint k = 1; k < p2->size(); k++) {
1300             p1->push(p2->at(k));
1301           }
1302           _packset.at_put(j, NULL);
1303           changed = true;
1304         }
1305       }
1306     }
1307   }
1308 
1309   // Split packs which have size greater then max vector size.
1310   for (int i = 0; i < _packset.length(); i++) {
1311     Node_List* p1 = _packset.at(i);
1312     if (p1 != NULL) {
1313       BasicType bt = velt_basic_type(p1->at(0));
1314       uint max_vlen = Matcher::max_vector_size(bt); // Max elements in vector
1315       assert(is_power_of_2(max_vlen), "sanity");
1316       uint psize = p1->size();
1317       if (!is_power_of_2(psize)) {
1318         // Skip pack which can't be vector.
1319         // case1: for(...) { a[i] = i; }    elements values are different (i+x)
1320         // case2: for(...) { a[i] = b[i+1]; }  can't align both, load and store
1321         _packset.at_put(i, NULL);
1322         continue;
1323       }
1324       if (psize > max_vlen) {
1325         Node_List* pack = new Node_List();
1326         for (uint j = 0; j < psize; j++) {
1327           pack->push(p1->at(j));
1328           if (pack->size() >= max_vlen) {
1329             assert(is_power_of_2(pack->size()), "sanity");
1330             _packset.append(pack);
1331             pack = new Node_List();
1332           }
1333         }
1334         _packset.at_put(i, NULL);
1335       }
1336     }
1337   }
1338 
1339   // Compress list.
1340   for (int i = _packset.length() - 1; i >= 0; i--) {
1341     Node_List* p1 = _packset.at(i);
1342     if (p1 == NULL) {
1343       _packset.remove_at(i);
1344     }
1345   }
1346 
1347 #ifndef PRODUCT
1348   if (TraceSuperWord) {
1349     tty->print_cr("\nAfter combine_packs");
1350     print_packset();
1351   }
1352 #endif
1353 }
1354 
1355 //-----------------------------construct_my_pack_map--------------------------
1356 // Construct the map from nodes to packs.  Only valid after the
1357 // point where a node is only in one pack (after combine_packs).
1358 void SuperWord::construct_my_pack_map() {
1359   Node_List* rslt = NULL;
1360   for (int i = 0; i < _packset.length(); i++) {
1361     Node_List* p = _packset.at(i);
1362     for (uint j = 0; j < p->size(); j++) {
1363       Node* s = p->at(j);
1364       assert(my_pack(s) == NULL, "only in one pack");
1365       set_my_pack(s, p);
1366     }
1367   }
1368 }
1369 
1370 //------------------------------filter_packs---------------------------
1371 // Remove packs that are not implemented or not profitable.
1372 void SuperWord::filter_packs() {
1373   // Remove packs that are not implemented
1374   for (int i = _packset.length() - 1; i >= 0; i--) {
1375     Node_List* pk = _packset.at(i);
1376     bool impl = implemented(pk);
1377     if (!impl) {
1378 #ifndef PRODUCT
1379       if (TraceSuperWord && Verbose) {
1380         tty->print_cr("Unimplemented");
1381         pk->at(0)->dump();
1382       }
1383 #endif
1384       remove_pack_at(i);
1385     }
1386     Node *n = pk->at(0);
1387     if (n->is_reduction()) {
1388       _num_reductions++;
1389     } else {
1390       _num_work_vecs++;
1391     }
1392   }
1393 
1394   // Remove packs that are not profitable
1395   bool changed;
1396   do {
1397     changed = false;
1398     for (int i = _packset.length() - 1; i >= 0; i--) {
1399       Node_List* pk = _packset.at(i);
1400       bool prof = profitable(pk);
1401       if (!prof) {
1402 #ifndef PRODUCT
1403         if (TraceSuperWord && Verbose) {
1404           tty->print_cr("Unprofitable");
1405           pk->at(0)->dump();
1406         }
1407 #endif
1408         remove_pack_at(i);
1409         changed = true;
1410       }
1411     }
1412   } while (changed);
1413 
1414 #ifndef PRODUCT
1415   if (TraceSuperWord) {
1416     tty->print_cr("\nAfter filter_packs");
1417     print_packset();
1418     tty->cr();
1419   }
1420 #endif
1421 }
1422 
1423 //------------------------------implemented---------------------------
1424 // Can code be generated for pack p?
1425 bool SuperWord::implemented(Node_List* p) {
1426   bool retValue = false;
1427   Node* p0 = p->at(0);
1428   if (p0 != NULL) {
1429     int opc = p0->Opcode();
1430     uint size = p->size();
1431     if (p0->is_reduction()) {
1432       const Type *arith_type = p0->bottom_type();
1433       // Length 2 reductions of INT/LONG do not offer performance benefits
1434       if (((arith_type->basic_type() == T_INT) || (arith_type->basic_type() == T_LONG)) && (size == 2)) {
1435         retValue = false;
1436       } else {
1437         retValue = ReductionNode::implemented(opc, size, arith_type->basic_type());
1438       }
1439     } else {
1440       retValue = VectorNode::implemented(opc, size, velt_basic_type(p0));
1441     }
1442   }
1443   return retValue;
1444 }
1445 
1446 //------------------------------same_inputs--------------------------
1447 // For pack p, are all idx operands the same?
1448 static bool same_inputs(Node_List* p, int idx) {
1449   Node* p0 = p->at(0);
1450   uint vlen = p->size();
1451   Node* p0_def = p0->in(idx);
1452   for (uint i = 1; i < vlen; i++) {
1453     Node* pi = p->at(i);
1454     Node* pi_def = pi->in(idx);
1455     if (p0_def != pi_def)
1456       return false;
1457   }
1458   return true;
1459 }
1460 
1461 //------------------------------profitable---------------------------
1462 // For pack p, are all operands and all uses (with in the block) vector?
1463 bool SuperWord::profitable(Node_List* p) {
1464   Node* p0 = p->at(0);
1465   uint start, end;
1466   VectorNode::vector_operands(p0, &start, &end);
1467 
1468   // Return false if some inputs are not vectors or vectors with different
1469   // size or alignment.
1470   // Also, for now, return false if not scalar promotion case when inputs are
1471   // the same. Later, implement PackNode and allow differing, non-vector inputs
1472   // (maybe just the ones from outside the block.)
1473   for (uint i = start; i < end; i++) {
1474     if (!is_vector_use(p0, i))
1475       return false;
1476   }
1477   // Check if reductions are connected
1478   if (p0->is_reduction()) {
1479     Node* second_in = p0->in(2);
1480     Node_List* second_pk = my_pack(second_in);
1481     if ((second_pk == NULL) || (_num_work_vecs == _num_reductions)) {
1482       // Remove reduction flag if no parent pack or if not enough work
1483       // to cover reduction expansion overhead
1484       p0->remove_flag(Node::Flag_is_reduction);
1485       return false;
1486     } else if (second_pk->size() != p->size()) {
1487       return false;
1488     }
1489   }
1490   if (VectorNode::is_shift(p0)) {
1491     // For now, return false if shift count is vector or not scalar promotion
1492     // case (different shift counts) because it is not supported yet.
1493     Node* cnt = p0->in(2);
1494     Node_List* cnt_pk = my_pack(cnt);
1495     if (cnt_pk != NULL)
1496       return false;
1497     if (!same_inputs(p, 2))
1498       return false;
1499   }
1500   if (!p0->is_Store()) {
1501     // For now, return false if not all uses are vector.
1502     // Later, implement ExtractNode and allow non-vector uses (maybe
1503     // just the ones outside the block.)
1504     for (uint i = 0; i < p->size(); i++) {
1505       Node* def = p->at(i);
1506       for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
1507         Node* use = def->fast_out(j);
1508         for (uint k = 0; k < use->req(); k++) {
1509           Node* n = use->in(k);
1510           if (def == n) {
1511             // reductions can be loop carried dependences
1512             if (def->is_reduction() && use->is_Phi())
1513               continue;
1514             if (!is_vector_use(use, k)) {
1515               return false;
1516             }
1517           }
1518         }
1519       }
1520     }
1521   }
1522   return true;
1523 }
1524 
1525 //------------------------------schedule---------------------------
1526 // Adjust the memory graph for the packed operations
1527 void SuperWord::schedule() {
1528 
1529   // Co-locate in the memory graph the members of each memory pack
1530   for (int i = 0; i < _packset.length(); i++) {
1531     co_locate_pack(_packset.at(i));
1532   }
1533 }
1534 
1535 //-------------------------------remove_and_insert-------------------
1536 // Remove "current" from its current position in the memory graph and insert
1537 // it after the appropriate insertion point (lip or uip).
1538 void SuperWord::remove_and_insert(MemNode *current, MemNode *prev, MemNode *lip,
1539                                   Node *uip, Unique_Node_List &sched_before) {
1540   Node* my_mem = current->in(MemNode::Memory);
1541   bool sched_up = sched_before.member(current);
1542 
1543   // remove current_store from its current position in the memmory graph
1544   for (DUIterator i = current->outs(); current->has_out(i); i++) {
1545     Node* use = current->out(i);
1546     if (use->is_Mem()) {
1547       assert(use->in(MemNode::Memory) == current, "must be");
1548       if (use == prev) { // connect prev to my_mem
1549           _igvn.replace_input_of(use, MemNode::Memory, my_mem);
1550           --i; //deleted this edge; rescan position
1551       } else if (sched_before.member(use)) {
1552         if (!sched_up) { // Will be moved together with current
1553           _igvn.replace_input_of(use, MemNode::Memory, uip);
1554           --i; //deleted this edge; rescan position
1555         }
1556       } else {
1557         if (sched_up) { // Will be moved together with current
1558           _igvn.replace_input_of(use, MemNode::Memory, lip);
1559           --i; //deleted this edge; rescan position
1560         }
1561       }
1562     }
1563   }
1564 
1565   Node *insert_pt =  sched_up ?  uip : lip;
1566 
1567   // all uses of insert_pt's memory state should use current's instead
1568   for (DUIterator i = insert_pt->outs(); insert_pt->has_out(i); i++) {
1569     Node* use = insert_pt->out(i);
1570     if (use->is_Mem()) {
1571       assert(use->in(MemNode::Memory) == insert_pt, "must be");
1572       _igvn.replace_input_of(use, MemNode::Memory, current);
1573       --i; //deleted this edge; rescan position
1574     } else if (!sched_up && use->is_Phi() && use->bottom_type() == Type::MEMORY) {
1575       uint pos; //lip (lower insert point) must be the last one in the memory slice
1576       for (pos=1; pos < use->req(); pos++) {
1577         if (use->in(pos) == insert_pt) break;
1578       }
1579       _igvn.replace_input_of(use, pos, current);
1580       --i;
1581     }
1582   }
1583 
1584   //connect current to insert_pt
1585   _igvn.replace_input_of(current, MemNode::Memory, insert_pt);
1586 }
1587 
1588 //------------------------------co_locate_pack----------------------------------
1589 // To schedule a store pack, we need to move any sandwiched memory ops either before
1590 // or after the pack, based upon dependence information:
1591 // (1) If any store in the pack depends on the sandwiched memory op, the
1592 //     sandwiched memory op must be scheduled BEFORE the pack;
1593 // (2) If a sandwiched memory op depends on any store in the pack, the
1594 //     sandwiched memory op must be scheduled AFTER the pack;
1595 // (3) If a sandwiched memory op (say, memA) depends on another sandwiched
1596 //     memory op (say memB), memB must be scheduled before memA. So, if memA is
1597 //     scheduled before the pack, memB must also be scheduled before the pack;
1598 // (4) If there is no dependence restriction for a sandwiched memory op, we simply
1599 //     schedule this store AFTER the pack
1600 // (5) We know there is no dependence cycle, so there in no other case;
1601 // (6) Finally, all memory ops in another single pack should be moved in the same direction.
1602 //
1603 // To schedule a load pack, we use the memory state of either the first or the last load in
1604 // the pack, based on the dependence constraint.
1605 void SuperWord::co_locate_pack(Node_List* pk) {
1606   if (pk->at(0)->is_Store()) {
1607     MemNode* first     = executed_first(pk)->as_Mem();
1608     MemNode* last      = executed_last(pk)->as_Mem();
1609     Unique_Node_List schedule_before_pack;
1610     Unique_Node_List memops;
1611 
1612     MemNode* current   = last->in(MemNode::Memory)->as_Mem();
1613     MemNode* previous  = last;
1614     while (true) {
1615       assert(in_bb(current), "stay in block");
1616       memops.push(previous);
1617       for (DUIterator i = current->outs(); current->has_out(i); i++) {
1618         Node* use = current->out(i);
1619         if (use->is_Mem() && use != previous)
1620           memops.push(use);
1621       }
1622       if (current == first) break;
1623       previous = current;
1624       current  = current->in(MemNode::Memory)->as_Mem();
1625     }
1626 
1627     // determine which memory operations should be scheduled before the pack
1628     for (uint i = 1; i < memops.size(); i++) {
1629       Node *s1 = memops.at(i);
1630       if (!in_pack(s1, pk) && !schedule_before_pack.member(s1)) {
1631         for (uint j = 0; j< i; j++) {
1632           Node *s2 = memops.at(j);
1633           if (!independent(s1, s2)) {
1634             if (in_pack(s2, pk) || schedule_before_pack.member(s2)) {
1635               schedule_before_pack.push(s1); // s1 must be scheduled before
1636               Node_List* mem_pk = my_pack(s1);
1637               if (mem_pk != NULL) {
1638                 for (uint ii = 0; ii < mem_pk->size(); ii++) {
1639                   Node* s = mem_pk->at(ii);  // follow partner
1640                   if (memops.member(s) && !schedule_before_pack.member(s))
1641                     schedule_before_pack.push(s);
1642                 }
1643               }
1644               break;
1645             }
1646           }
1647         }
1648       }
1649     }
1650 
1651     Node*    upper_insert_pt = first->in(MemNode::Memory);
1652     // Following code moves loads connected to upper_insert_pt below aliased stores.
1653     // Collect such loads here and reconnect them back to upper_insert_pt later.
1654     memops.clear();
1655     for (DUIterator i = upper_insert_pt->outs(); upper_insert_pt->has_out(i); i++) {
1656       Node* use = upper_insert_pt->out(i);
1657       if (use->is_Mem() && !use->is_Store()) {
1658         memops.push(use);
1659       }
1660     }
1661 
1662     MemNode* lower_insert_pt = last;
1663     previous                 = last; //previous store in pk
1664     current                  = last->in(MemNode::Memory)->as_Mem();
1665 
1666     // start scheduling from "last" to "first"
1667     while (true) {
1668       assert(in_bb(current), "stay in block");
1669       assert(in_pack(previous, pk), "previous stays in pack");
1670       Node* my_mem = current->in(MemNode::Memory);
1671 
1672       if (in_pack(current, pk)) {
1673         // Forward users of my memory state (except "previous) to my input memory state
1674         for (DUIterator i = current->outs(); current->has_out(i); i++) {
1675           Node* use = current->out(i);
1676           if (use->is_Mem() && use != previous) {
1677             assert(use->in(MemNode::Memory) == current, "must be");
1678             if (schedule_before_pack.member(use)) {
1679               _igvn.replace_input_of(use, MemNode::Memory, upper_insert_pt);
1680             } else {
1681               _igvn.replace_input_of(use, MemNode::Memory, lower_insert_pt);
1682             }
1683             --i; // deleted this edge; rescan position
1684           }
1685         }
1686         previous = current;
1687       } else { // !in_pack(current, pk) ==> a sandwiched store
1688         remove_and_insert(current, previous, lower_insert_pt, upper_insert_pt, schedule_before_pack);
1689       }
1690 
1691       if (current == first) break;
1692       current = my_mem->as_Mem();
1693     } // end while
1694 
1695     // Reconnect loads back to upper_insert_pt.
1696     for (uint i = 0; i < memops.size(); i++) {
1697       Node *ld = memops.at(i);
1698       if (ld->in(MemNode::Memory) != upper_insert_pt) {
1699         _igvn.replace_input_of(ld, MemNode::Memory, upper_insert_pt);
1700       }
1701     }
1702   } else if (pk->at(0)->is_Load()) { //load
1703     // all loads in the pack should have the same memory state. By default,
1704     // we use the memory state of the last load. However, if any load could
1705     // not be moved down due to the dependence constraint, we use the memory
1706     // state of the first load.
1707     Node* last_mem  = executed_last(pk)->in(MemNode::Memory);
1708     Node* first_mem = executed_first(pk)->in(MemNode::Memory);
1709     bool schedule_last = true;
1710     for (uint i = 0; i < pk->size(); i++) {
1711       Node* ld = pk->at(i);
1712       for (Node* current = last_mem; current != ld->in(MemNode::Memory);
1713            current=current->in(MemNode::Memory)) {
1714         assert(current != first_mem, "corrupted memory graph");
1715         if(current->is_Mem() && !independent(current, ld)){
1716           schedule_last = false; // a later store depends on this load
1717           break;
1718         }
1719       }
1720     }
1721 
1722     Node* mem_input = schedule_last ? last_mem : first_mem;
1723     _igvn.hash_delete(mem_input);
1724     // Give each load the same memory state
1725     for (uint i = 0; i < pk->size(); i++) {
1726       LoadNode* ld = pk->at(i)->as_Load();
1727       _igvn.replace_input_of(ld, MemNode::Memory, mem_input);
1728     }
1729   }
1730 }
1731 
1732 //------------------------------output---------------------------
1733 // Convert packs into vector node operations
1734 void SuperWord::output() {
1735   if (_packset.length() == 0) return;
1736 
1737 #ifndef PRODUCT
1738   if (TraceLoopOpts) {
1739     tty->print("SuperWord    ");
1740     lpt()->dump_head();
1741   }
1742 #endif
1743 
1744   // MUST ENSURE main loop's initial value is properly aligned:
1745   //  (iv_initial_value + min_iv_offset) % vector_width_in_bytes() == 0
1746 
1747   align_initial_loop_index(align_to_ref());
1748 
1749   // Insert extract (unpack) operations for scalar uses
1750   for (int i = 0; i < _packset.length(); i++) {
1751     insert_extracts(_packset.at(i));
1752   }
1753 
1754   Compile* C = _phase->C;
1755   CountedLoopNode *cl = lpt()->_head->as_CountedLoop();
1756   uint max_vlen_in_bytes = 0;
1757   uint max_vlen = 0;
1758   for (int i = 0; i < _block.length(); i++) {
1759     Node* n = _block.at(i);
1760     Node_List* p = my_pack(n);
1761     if (p && n == executed_last(p)) {
1762       uint vlen = p->size();
1763       uint vlen_in_bytes = 0;
1764       Node* vn = NULL;
1765       Node* low_adr = p->at(0);
1766       Node* first   = executed_first(p);
1767       int   opc = n->Opcode();
1768       if (n->is_Load()) {
1769         Node* ctl = n->in(MemNode::Control);
1770         Node* mem = first->in(MemNode::Memory);
1771         SWPointer p1(n->as_Mem(), this, NULL, false);
1772         // Identify the memory dependency for the new loadVector node by
1773         // walking up through memory chain.
1774         // This is done to give flexibility to the new loadVector node so that
1775         // it can move above independent storeVector nodes.
1776         while (mem->is_StoreVector()) {
1777           SWPointer p2(mem->as_Mem(), this, NULL, false);
1778           int cmp = p1.cmp(p2);
1779           if (SWPointer::not_equal(cmp) || !SWPointer::comparable(cmp)) {
1780             mem = mem->in(MemNode::Memory);
1781           } else {
1782             break; // dependent memory
1783           }
1784         }
1785         Node* adr = low_adr->in(MemNode::Address);
1786         const TypePtr* atyp = n->adr_type();
1787         vn = LoadVectorNode::make(opc, ctl, mem, adr, atyp, vlen, velt_basic_type(n), control_dependency(p));
1788         vlen_in_bytes = vn->as_LoadVector()->memory_size();
1789       } else if (n->is_Store()) {
1790         // Promote value to be stored to vector
1791         Node* val = vector_opd(p, MemNode::ValueIn);
1792         Node* ctl = n->in(MemNode::Control);
1793         Node* mem = first->in(MemNode::Memory);
1794         Node* adr = low_adr->in(MemNode::Address);
1795         const TypePtr* atyp = n->adr_type();
1796         vn = StoreVectorNode::make(opc, ctl, mem, adr, atyp, val, vlen);
1797         vlen_in_bytes = vn->as_StoreVector()->memory_size();
1798       } else if (n->req() == 3) {
1799         // Promote operands to vector
1800         Node* in1 = NULL;
1801         bool node_isa_reduction = n->is_reduction();
1802         if (node_isa_reduction) {
1803           // the input to the first reduction operation is retained
1804           in1 = low_adr->in(1);
1805         } else {
1806           in1 = vector_opd(p, 1);
1807         }
1808         Node* in2 = vector_opd(p, 2);
1809         if (VectorNode::is_invariant_vector(in1) && (node_isa_reduction == false) && (n->is_Add() || n->is_Mul())) {
1810           // Move invariant vector input into second position to avoid register spilling.
1811           Node* tmp = in1;
1812           in1 = in2;
1813           in2 = tmp;
1814         }
1815         if (node_isa_reduction) {
1816           const Type *arith_type = n->bottom_type();
1817           vn = ReductionNode::make(opc, NULL, in1, in2, arith_type->basic_type());
1818           if (in2->is_Load()) {
1819             vlen_in_bytes = in2->as_LoadVector()->memory_size();
1820           } else {
1821             vlen_in_bytes = in2->as_Vector()->length_in_bytes();
1822           }
1823         } else {
1824           vn = VectorNode::make(opc, in1, in2, vlen, velt_basic_type(n));
1825           vlen_in_bytes = vn->as_Vector()->length_in_bytes();
1826         }
1827       } else {
1828         ShouldNotReachHere();
1829       }
1830       assert(vn != NULL, "sanity");
1831       _igvn.register_new_node_with_optimizer(vn);
1832       _phase->set_ctrl(vn, _phase->get_ctrl(p->at(0)));
1833       for (uint j = 0; j < p->size(); j++) {
1834         Node* pm = p->at(j);
1835         _igvn.replace_node(pm, vn);
1836       }
1837       _igvn._worklist.push(vn);
1838 
1839       if (vlen_in_bytes > max_vlen_in_bytes) {
1840         max_vlen = vlen;
1841         max_vlen_in_bytes = vlen_in_bytes;
1842       }
1843 #ifdef ASSERT
1844       if (TraceNewVectors) {
1845         tty->print("new Vector node: ");
1846         vn->dump();
1847       }
1848 #endif
1849     }
1850   }
1851   C->set_max_vector_size(max_vlen_in_bytes);
1852   if (SuperWordLoopUnrollAnalysis) {
1853     if (cl->has_passed_slp()) {
1854       int slp_max_unroll_factor = cl->slp_max_unroll();
1855       if (slp_max_unroll_factor == max_vlen) {
1856         NOT_PRODUCT(if (TraceSuperWordLoopUnrollAnalysis) tty->print_cr("vector loop(unroll=%d, len=%d)\n", max_vlen, max_vlen_in_bytes*BitsPerByte));
1857         // For atomic unrolled loops which are vector mapped, instigate more unrolling.
1858         cl->set_notpassed_slp();
1859         C->set_major_progress();
1860         cl->mark_no_slp();
1861       }
1862     }
1863   }
1864 }
1865 
1866 //------------------------------vector_opd---------------------------
1867 // Create a vector operand for the nodes in pack p for operand: in(opd_idx)
1868 Node* SuperWord::vector_opd(Node_List* p, int opd_idx) {
1869   Node* p0 = p->at(0);
1870   uint vlen = p->size();
1871   Node* opd = p0->in(opd_idx);
1872 
1873   if (same_inputs(p, opd_idx)) {
1874     if (opd->is_Vector() || opd->is_LoadVector()) {
1875       assert(((opd_idx != 2) || !VectorNode::is_shift(p0)), "shift's count can't be vector");
1876       return opd; // input is matching vector
1877     }
1878     if ((opd_idx == 2) && VectorNode::is_shift(p0)) {
1879       Compile* C = _phase->C;
1880       Node* cnt = opd;
1881       // Vector instructions do not mask shift count, do it here.
1882       juint mask = (p0->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
1883       const TypeInt* t = opd->find_int_type();
1884       if (t != NULL && t->is_con()) {
1885         juint shift = t->get_con();
1886         if (shift > mask) { // Unsigned cmp
1887           cnt = ConNode::make(TypeInt::make(shift & mask));
1888         }
1889       } else {
1890         if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
1891           cnt = ConNode::make(TypeInt::make(mask));
1892           _igvn.register_new_node_with_optimizer(cnt);
1893           cnt = new AndINode(opd, cnt);
1894           _igvn.register_new_node_with_optimizer(cnt);
1895           _phase->set_ctrl(cnt, _phase->get_ctrl(opd));
1896         }
1897         assert(opd->bottom_type()->isa_int(), "int type only");
1898         // Move non constant shift count into vector register.
1899         cnt = VectorNode::shift_count(p0, cnt, vlen, velt_basic_type(p0));
1900       }
1901       if (cnt != opd) {
1902         _igvn.register_new_node_with_optimizer(cnt);
1903         _phase->set_ctrl(cnt, _phase->get_ctrl(opd));
1904       }
1905       return cnt;
1906     }
1907     assert(!opd->is_StoreVector(), "such vector is not expected here");
1908     // Convert scalar input to vector with the same number of elements as
1909     // p0's vector. Use p0's type because size of operand's container in
1910     // vector should match p0's size regardless operand's size.
1911     const Type* p0_t = velt_type(p0);
1912     VectorNode* vn = VectorNode::scalar2vector(opd, vlen, p0_t);
1913 
1914     _igvn.register_new_node_with_optimizer(vn);
1915     _phase->set_ctrl(vn, _phase->get_ctrl(opd));
1916 #ifdef ASSERT
1917     if (TraceNewVectors) {
1918       tty->print("new Vector node: ");
1919       vn->dump();
1920     }
1921 #endif
1922     return vn;
1923   }
1924 
1925   // Insert pack operation
1926   BasicType bt = velt_basic_type(p0);
1927   PackNode* pk = PackNode::make(opd, vlen, bt);
1928   DEBUG_ONLY( const BasicType opd_bt = opd->bottom_type()->basic_type(); )
1929 
1930   for (uint i = 1; i < vlen; i++) {
1931     Node* pi = p->at(i);
1932     Node* in = pi->in(opd_idx);
1933     assert(my_pack(in) == NULL, "Should already have been unpacked");
1934     assert(opd_bt == in->bottom_type()->basic_type(), "all same type");
1935     pk->add_opd(in);
1936   }
1937   _igvn.register_new_node_with_optimizer(pk);
1938   _phase->set_ctrl(pk, _phase->get_ctrl(opd));
1939 #ifdef ASSERT
1940   if (TraceNewVectors) {
1941     tty->print("new Vector node: ");
1942     pk->dump();
1943   }
1944 #endif
1945   return pk;
1946 }
1947 
1948 //------------------------------insert_extracts---------------------------
1949 // If a use of pack p is not a vector use, then replace the
1950 // use with an extract operation.
1951 void SuperWord::insert_extracts(Node_List* p) {
1952   if (p->at(0)->is_Store()) return;
1953   assert(_n_idx_list.is_empty(), "empty (node,index) list");
1954 
1955   // Inspect each use of each pack member.  For each use that is
1956   // not a vector use, replace the use with an extract operation.
1957 
1958   for (uint i = 0; i < p->size(); i++) {
1959     Node* def = p->at(i);
1960     for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
1961       Node* use = def->fast_out(j);
1962       for (uint k = 0; k < use->req(); k++) {
1963         Node* n = use->in(k);
1964         if (def == n) {
1965           if (!is_vector_use(use, k)) {
1966             _n_idx_list.push(use, k);
1967           }
1968         }
1969       }
1970     }
1971   }
1972 
1973   while (_n_idx_list.is_nonempty()) {
1974     Node* use = _n_idx_list.node();
1975     int   idx = _n_idx_list.index();
1976     _n_idx_list.pop();
1977     Node* def = use->in(idx);
1978 
1979     if (def->is_reduction()) continue;
1980 
1981     // Insert extract operation
1982     _igvn.hash_delete(def);
1983     int def_pos = alignment(def) / data_size(def);
1984 
1985     Node* ex = ExtractNode::make(def, def_pos, velt_basic_type(def));
1986     _igvn.register_new_node_with_optimizer(ex);
1987     _phase->set_ctrl(ex, _phase->get_ctrl(def));
1988     _igvn.replace_input_of(use, idx, ex);
1989     _igvn._worklist.push(def);
1990 
1991     bb_insert_after(ex, bb_idx(def));
1992     set_velt_type(ex, velt_type(def));
1993   }
1994 }
1995 
1996 //------------------------------is_vector_use---------------------------
1997 // Is use->in(u_idx) a vector use?
1998 bool SuperWord::is_vector_use(Node* use, int u_idx) {
1999   Node_List* u_pk = my_pack(use);
2000   if (u_pk == NULL) return false;
2001   if (use->is_reduction()) return true;
2002   Node* def = use->in(u_idx);
2003   Node_List* d_pk = my_pack(def);
2004   if (d_pk == NULL) {
2005     // check for scalar promotion
2006     Node* n = u_pk->at(0)->in(u_idx);
2007     for (uint i = 1; i < u_pk->size(); i++) {
2008       if (u_pk->at(i)->in(u_idx) != n) return false;
2009     }
2010     return true;
2011   }
2012   if (u_pk->size() != d_pk->size())
2013     return false;
2014   for (uint i = 0; i < u_pk->size(); i++) {
2015     Node* ui = u_pk->at(i);
2016     Node* di = d_pk->at(i);
2017     if (ui->in(u_idx) != di || alignment(ui) != alignment(di))
2018       return false;
2019   }
2020   return true;
2021 }
2022 
2023 //------------------------------construct_bb---------------------------
2024 // Construct reverse postorder list of block members
2025 bool SuperWord::construct_bb() {
2026   Node* entry = bb();
2027 
2028   assert(_stk.length() == 0,            "stk is empty");
2029   assert(_block.length() == 0,          "block is empty");
2030   assert(_data_entry.length() == 0,     "data_entry is empty");
2031   assert(_mem_slice_head.length() == 0, "mem_slice_head is empty");
2032   assert(_mem_slice_tail.length() == 0, "mem_slice_tail is empty");
2033 
2034   // Find non-control nodes with no inputs from within block,
2035   // create a temporary map from node _idx to bb_idx for use
2036   // by the visited and post_visited sets,
2037   // and count number of nodes in block.
2038   int bb_ct = 0;
2039   for (uint i = 0; i < lpt()->_body.size(); i++) {
2040     Node *n = lpt()->_body.at(i);
2041     set_bb_idx(n, i); // Create a temporary map
2042     if (in_bb(n)) {
2043       if (n->is_LoadStore() || n->is_MergeMem() ||
2044           (n->is_Proj() && !n->as_Proj()->is_CFG())) {
2045         // Bailout if the loop has LoadStore, MergeMem or data Proj
2046         // nodes. Superword optimization does not work with them.
2047         return false;
2048       }
2049       bb_ct++;
2050       if (!n->is_CFG()) {
2051         bool found = false;
2052         for (uint j = 0; j < n->req(); j++) {
2053           Node* def = n->in(j);
2054           if (def && in_bb(def)) {
2055             found = true;
2056             break;
2057           }
2058         }
2059         if (!found) {
2060           assert(n != entry, "can't be entry");
2061           _data_entry.push(n);
2062         }
2063       }
2064     }
2065   }
2066 
2067   // Find memory slices (head and tail)
2068   for (DUIterator_Fast imax, i = lp()->fast_outs(imax); i < imax; i++) {
2069     Node *n = lp()->fast_out(i);
2070     if (in_bb(n) && (n->is_Phi() && n->bottom_type() == Type::MEMORY)) {
2071       Node* n_tail  = n->in(LoopNode::LoopBackControl);
2072       if (n_tail != n->in(LoopNode::EntryControl)) {
2073         if (!n_tail->is_Mem()) {
2074           assert(n_tail->is_Mem(), err_msg_res("unexpected node for memory slice: %s", n_tail->Name()));
2075           return false; // Bailout
2076         }
2077         _mem_slice_head.push(n);
2078         _mem_slice_tail.push(n_tail);
2079       }
2080     }
2081   }
2082 
2083   // Create an RPO list of nodes in block
2084 
2085   visited_clear();
2086   post_visited_clear();
2087 
2088   // Push all non-control nodes with no inputs from within block, then control entry
2089   for (int j = 0; j < _data_entry.length(); j++) {
2090     Node* n = _data_entry.at(j);
2091     visited_set(n);
2092     _stk.push(n);
2093   }
2094   visited_set(entry);
2095   _stk.push(entry);
2096 
2097   // Do a depth first walk over out edges
2098   int rpo_idx = bb_ct - 1;
2099   int size;
2100   int reduction_uses = 0;
2101   while ((size = _stk.length()) > 0) {
2102     Node* n = _stk.top(); // Leave node on stack
2103     if (!visited_test_set(n)) {
2104       // forward arc in graph
2105     } else if (!post_visited_test(n)) {
2106       // cross or back arc
2107       for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2108         Node *use = n->fast_out(i);
2109         if (in_bb(use) && !visited_test(use) &&
2110             // Don't go around backedge
2111             (!use->is_Phi() || n == entry)) {
2112           if (use->is_reduction()) {
2113             // First see if we can map the reduction on the given system we are on, then
2114             // make a data entry operation for each reduction we see.
2115             BasicType bt = use->bottom_type()->basic_type();
2116             if (ReductionNode::implemented(use->Opcode(), Matcher::min_vector_size(bt), bt)) {
2117               reduction_uses++;
2118             }
2119           }
2120           _stk.push(use);
2121         }
2122       }
2123       if (_stk.length() == size) {
2124         // There were no additional uses, post visit node now
2125         _stk.pop(); // Remove node from stack
2126         assert(rpo_idx >= 0, "");
2127         _block.at_put_grow(rpo_idx, n);
2128         rpo_idx--;
2129         post_visited_set(n);
2130         assert(rpo_idx >= 0 || _stk.is_empty(), "");
2131       }
2132     } else {
2133       _stk.pop(); // Remove post-visited node from stack
2134     }
2135   }
2136 
2137   // Create real map of block indices for nodes
2138   for (int j = 0; j < _block.length(); j++) {
2139     Node* n = _block.at(j);
2140     set_bb_idx(n, j);
2141   }
2142 
2143   // Ensure extra info is allocated.
2144   initialize_bb();
2145 
2146 #ifndef PRODUCT
2147   if (TraceSuperWord) {
2148     print_bb();
2149     tty->print_cr("\ndata entry nodes: %s", _data_entry.length() > 0 ? "" : "NONE");
2150     for (int m = 0; m < _data_entry.length(); m++) {
2151       tty->print("%3d ", m);
2152       _data_entry.at(m)->dump();
2153     }
2154     tty->print_cr("\nmemory slices: %s", _mem_slice_head.length() > 0 ? "" : "NONE");
2155     for (int m = 0; m < _mem_slice_head.length(); m++) {
2156       tty->print("%3d ", m); _mem_slice_head.at(m)->dump();
2157       tty->print("    ");    _mem_slice_tail.at(m)->dump();
2158     }
2159   }
2160 #endif
2161   assert(rpo_idx == -1 && bb_ct == _block.length(), "all block members found");
2162   return (_mem_slice_head.length() > 0) || (reduction_uses > 0) || (_data_entry.length() > 0);
2163 }
2164 
2165 //------------------------------initialize_bb---------------------------
2166 // Initialize per node info
2167 void SuperWord::initialize_bb() {
2168   Node* last = _block.at(_block.length() - 1);
2169   grow_node_info(bb_idx(last));
2170 }
2171 
2172 //------------------------------bb_insert_after---------------------------
2173 // Insert n into block after pos
2174 void SuperWord::bb_insert_after(Node* n, int pos) {
2175   int n_pos = pos + 1;
2176   // Make room
2177   for (int i = _block.length() - 1; i >= n_pos; i--) {
2178     _block.at_put_grow(i+1, _block.at(i));
2179   }
2180   for (int j = _node_info.length() - 1; j >= n_pos; j--) {
2181     _node_info.at_put_grow(j+1, _node_info.at(j));
2182   }
2183   // Set value
2184   _block.at_put_grow(n_pos, n);
2185   _node_info.at_put_grow(n_pos, SWNodeInfo::initial);
2186   // Adjust map from node->_idx to _block index
2187   for (int i = n_pos; i < _block.length(); i++) {
2188     set_bb_idx(_block.at(i), i);
2189   }
2190 }
2191 
2192 //------------------------------compute_max_depth---------------------------
2193 // Compute max depth for expressions from beginning of block
2194 // Use to prune search paths during test for independence.
2195 void SuperWord::compute_max_depth() {
2196   int ct = 0;
2197   bool again;
2198   do {
2199     again = false;
2200     for (int i = 0; i < _block.length(); i++) {
2201       Node* n = _block.at(i);
2202       if (!n->is_Phi()) {
2203         int d_orig = depth(n);
2204         int d_in   = 0;
2205         for (DepPreds preds(n, _dg); !preds.done(); preds.next()) {
2206           Node* pred = preds.current();
2207           if (in_bb(pred)) {
2208             d_in = MAX2(d_in, depth(pred));
2209           }
2210         }
2211         if (d_in + 1 != d_orig) {
2212           set_depth(n, d_in + 1);
2213           again = true;
2214         }
2215       }
2216     }
2217     ct++;
2218   } while (again);
2219 #ifndef PRODUCT
2220   if (TraceSuperWord && Verbose)
2221     tty->print_cr("compute_max_depth iterated: %d times", ct);
2222 #endif
2223 }
2224 
2225 //-------------------------compute_vector_element_type-----------------------
2226 // Compute necessary vector element type for expressions
2227 // This propagates backwards a narrower integer type when the
2228 // upper bits of the value are not needed.
2229 // Example:  char a,b,c;  a = b + c;
2230 // Normally the type of the add is integer, but for packed character
2231 // operations the type of the add needs to be char.
2232 void SuperWord::compute_vector_element_type() {
2233 #ifndef PRODUCT
2234   if (TraceSuperWord && Verbose)
2235     tty->print_cr("\ncompute_velt_type:");
2236 #endif
2237 
2238   // Initial type
2239   for (int i = 0; i < _block.length(); i++) {
2240     Node* n = _block.at(i);
2241     set_velt_type(n, container_type(n));
2242   }
2243 
2244   // Propagate integer narrowed type backwards through operations
2245   // that don't depend on higher order bits
2246   for (int i = _block.length() - 1; i >= 0; i--) {
2247     Node* n = _block.at(i);
2248     // Only integer types need be examined
2249     const Type* vtn = velt_type(n);
2250     if (vtn->basic_type() == T_INT) {
2251       uint start, end;
2252       VectorNode::vector_operands(n, &start, &end);
2253 
2254       for (uint j = start; j < end; j++) {
2255         Node* in  = n->in(j);
2256         // Don't propagate through a memory
2257         if (!in->is_Mem() && in_bb(in) && velt_type(in)->basic_type() == T_INT &&
2258             data_size(n) < data_size(in)) {
2259           bool same_type = true;
2260           for (DUIterator_Fast kmax, k = in->fast_outs(kmax); k < kmax; k++) {
2261             Node *use = in->fast_out(k);
2262             if (!in_bb(use) || !same_velt_type(use, n)) {
2263               same_type = false;
2264               break;
2265             }
2266           }
2267           if (same_type) {
2268             // For right shifts of small integer types (bool, byte, char, short)
2269             // we need precise information about sign-ness. Only Load nodes have
2270             // this information because Store nodes are the same for signed and
2271             // unsigned values. And any arithmetic operation after a load may
2272             // expand a value to signed Int so such right shifts can't be used
2273             // because vector elements do not have upper bits of Int.
2274             const Type* vt = vtn;
2275             if (VectorNode::is_shift(in)) {
2276               Node* load = in->in(1);
2277               if (load->is_Load() && in_bb(load) && (velt_type(load)->basic_type() == T_INT)) {
2278                 vt = velt_type(load);
2279               } else if (in->Opcode() != Op_LShiftI) {
2280                 // Widen type to Int to avoid creation of right shift vector
2281                 // (align + data_size(s1) check in stmts_can_pack() will fail).
2282                 // Note, left shifts work regardless type.
2283                 vt = TypeInt::INT;
2284               }
2285             }
2286             set_velt_type(in, vt);
2287           }
2288         }
2289       }
2290     }
2291   }
2292 #ifndef PRODUCT
2293   if (TraceSuperWord && Verbose) {
2294     for (int i = 0; i < _block.length(); i++) {
2295       Node* n = _block.at(i);
2296       velt_type(n)->dump();
2297       tty->print("\t");
2298       n->dump();
2299     }
2300   }
2301 #endif
2302 }
2303 
2304 //------------------------------memory_alignment---------------------------
2305 // Alignment within a vector memory reference
2306 int SuperWord::memory_alignment(MemNode* s, int iv_adjust) {
2307   SWPointer p(s, this, NULL, false);
2308   if (!p.valid()) {
2309     return bottom_align;
2310   }
2311   int vw = vector_width_in_bytes(s);
2312   if (vw < 2) {
2313     return bottom_align; // No vectors for this type
2314   }
2315   int offset  = p.offset_in_bytes();
2316   offset     += iv_adjust*p.memory_size();
2317   int off_rem = offset % vw;
2318   int off_mod = off_rem >= 0 ? off_rem : off_rem + vw;
2319   return off_mod;
2320 }
2321 
2322 //---------------------------container_type---------------------------
2323 // Smallest type containing range of values
2324 const Type* SuperWord::container_type(Node* n) {
2325   if (n->is_Mem()) {
2326     BasicType bt = n->as_Mem()->memory_type();
2327     if (n->is_Store() && (bt == T_CHAR)) {
2328       // Use T_SHORT type instead of T_CHAR for stored values because any
2329       // preceding arithmetic operation extends values to signed Int.
2330       bt = T_SHORT;
2331     }
2332     if (n->Opcode() == Op_LoadUB) {
2333       // Adjust type for unsigned byte loads, it is important for right shifts.
2334       // T_BOOLEAN is used because there is no basic type representing type
2335       // TypeInt::UBYTE. Use of T_BOOLEAN for vectors is fine because only
2336       // size (one byte) and sign is important.
2337       bt = T_BOOLEAN;
2338     }
2339     return Type::get_const_basic_type(bt);
2340   }
2341   const Type* t = _igvn.type(n);
2342   if (t->basic_type() == T_INT) {
2343     // A narrow type of arithmetic operations will be determined by
2344     // propagating the type of memory operations.
2345     return TypeInt::INT;
2346   }
2347   return t;
2348 }
2349 
2350 bool SuperWord::same_velt_type(Node* n1, Node* n2) {
2351   const Type* vt1 = velt_type(n1);
2352   const Type* vt2 = velt_type(n2);
2353   if (vt1->basic_type() == T_INT && vt2->basic_type() == T_INT) {
2354     // Compare vectors element sizes for integer types.
2355     return data_size(n1) == data_size(n2);
2356   }
2357   return vt1 == vt2;
2358 }
2359 
2360 //------------------------------in_packset---------------------------
2361 // Are s1 and s2 in a pack pair and ordered as s1,s2?
2362 bool SuperWord::in_packset(Node* s1, Node* s2) {
2363   for (int i = 0; i < _packset.length(); i++) {
2364     Node_List* p = _packset.at(i);
2365     assert(p->size() == 2, "must be");
2366     if (p->at(0) == s1 && p->at(p->size()-1) == s2) {
2367       return true;
2368     }
2369   }
2370   return false;
2371 }
2372 
2373 //------------------------------in_pack---------------------------
2374 // Is s in pack p?
2375 Node_List* SuperWord::in_pack(Node* s, Node_List* p) {
2376   for (uint i = 0; i < p->size(); i++) {
2377     if (p->at(i) == s) {
2378       return p;
2379     }
2380   }
2381   return NULL;
2382 }
2383 
2384 //------------------------------remove_pack_at---------------------------
2385 // Remove the pack at position pos in the packset
2386 void SuperWord::remove_pack_at(int pos) {
2387   Node_List* p = _packset.at(pos);
2388   for (uint i = 0; i < p->size(); i++) {
2389     Node* s = p->at(i);
2390     set_my_pack(s, NULL);
2391   }
2392   _packset.remove_at(pos);
2393 }
2394 
2395 void SuperWord::packset_sort(int n) {
2396   // simple bubble sort so that we capitalize with O(n) when its already sorted
2397   while (n != 0) {
2398     bool swapped = false;
2399     for (int i = 1; i < n; i++) {
2400       Node_List* q_low = _packset.at(i-1);
2401       Node_List* q_i = _packset.at(i);
2402 
2403       // only swap when we find something to swap
2404       if (alignment(q_low->at(0)) > alignment(q_i->at(0))) {
2405         Node_List* t = q_i;
2406         *(_packset.adr_at(i)) = q_low;
2407         *(_packset.adr_at(i-1)) = q_i;
2408         swapped = true;
2409       }
2410     }
2411     if (swapped == false) break;
2412     n--;
2413   }
2414 }
2415 
2416 //------------------------------executed_first---------------------------
2417 // Return the node executed first in pack p.  Uses the RPO block list
2418 // to determine order.
2419 Node* SuperWord::executed_first(Node_List* p) {
2420   Node* n = p->at(0);
2421   int n_rpo = bb_idx(n);
2422   for (uint i = 1; i < p->size(); i++) {
2423     Node* s = p->at(i);
2424     int s_rpo = bb_idx(s);
2425     if (s_rpo < n_rpo) {
2426       n = s;
2427       n_rpo = s_rpo;
2428     }
2429   }
2430   return n;
2431 }
2432 
2433 //------------------------------executed_last---------------------------
2434 // Return the node executed last in pack p.
2435 Node* SuperWord::executed_last(Node_List* p) {
2436   Node* n = p->at(0);
2437   int n_rpo = bb_idx(n);
2438   for (uint i = 1; i < p->size(); i++) {
2439     Node* s = p->at(i);
2440     int s_rpo = bb_idx(s);
2441     if (s_rpo > n_rpo) {
2442       n = s;
2443       n_rpo = s_rpo;
2444     }
2445   }
2446   return n;
2447 }
2448 
2449 LoadNode::ControlDependency SuperWord::control_dependency(Node_List* p) {
2450   LoadNode::ControlDependency dep = LoadNode::DependsOnlyOnTest;
2451   for (uint i = 0; i < p->size(); i++) {
2452     Node* n = p->at(i);
2453     assert(n->is_Load(), "only meaningful for loads");
2454     if (!n->depends_only_on_test()) {
2455       dep = LoadNode::Pinned;
2456     }
2457   }
2458   return dep;
2459 }
2460 
2461 
2462 //----------------------------align_initial_loop_index---------------------------
2463 // Adjust pre-loop limit so that in main loop, a load/store reference
2464 // to align_to_ref will be a position zero in the vector.
2465 //   (iv + k) mod vector_align == 0
2466 void SuperWord::align_initial_loop_index(MemNode* align_to_ref) {
2467   CountedLoopNode *main_head = lp()->as_CountedLoop();
2468   assert(main_head->is_main_loop(), "");
2469   CountedLoopEndNode* pre_end = get_pre_loop_end(main_head);
2470   assert(pre_end != NULL, "we must have a correct pre-loop");
2471   Node *pre_opaq1 = pre_end->limit();
2472   assert(pre_opaq1->Opcode() == Op_Opaque1, "");
2473   Opaque1Node *pre_opaq = (Opaque1Node*)pre_opaq1;
2474   Node *lim0 = pre_opaq->in(1);
2475 
2476   // Where we put new limit calculations
2477   Node *pre_ctrl = pre_end->loopnode()->in(LoopNode::EntryControl);
2478 
2479   // Ensure the original loop limit is available from the
2480   // pre-loop Opaque1 node.
2481   Node *orig_limit = pre_opaq->original_loop_limit();
2482   assert(orig_limit != NULL && _igvn.type(orig_limit) != Type::TOP, "");
2483 
2484   SWPointer align_to_ref_p(align_to_ref, this, NULL, false);
2485   assert(align_to_ref_p.valid(), "sanity");
2486 
2487   // Given:
2488   //     lim0 == original pre loop limit
2489   //     V == v_align (power of 2)
2490   //     invar == extra invariant piece of the address expression
2491   //     e == offset [ +/- invar ]
2492   //
2493   // When reassociating expressions involving '%' the basic rules are:
2494   //     (a - b) % k == 0   =>  a % k == b % k
2495   // and:
2496   //     (a + b) % k == 0   =>  a % k == (k - b) % k
2497   //
2498   // For stride > 0 && scale > 0,
2499   //   Derive the new pre-loop limit "lim" such that the two constraints:
2500   //     (1) lim = lim0 + N           (where N is some positive integer < V)
2501   //     (2) (e + lim) % V == 0
2502   //   are true.
2503   //
2504   //   Substituting (1) into (2),
2505   //     (e + lim0 + N) % V == 0
2506   //   solve for N:
2507   //     N = (V - (e + lim0)) % V
2508   //   substitute back into (1), so that new limit
2509   //     lim = lim0 + (V - (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 = (e + lim0) % V
2527   //     lim = lim0 - (e + lim0) % V
2528   //
2529   // For stride < 0 && scale < 0
2530   //   Constraints:
2531   //     lim = lim0 - N
2532   //     (e - lim) % V == 0
2533   //   Solving for lim:
2534   //     (e - lim0 + N) % V == 0
2535   //     N = (V - (e - lim0)) % V
2536   //     lim = lim0 - (V - (e - lim0)) % V
2537 
2538   int vw = vector_width_in_bytes(align_to_ref);
2539   int stride   = iv_stride();
2540   int scale    = align_to_ref_p.scale_in_bytes();
2541   int elt_size = align_to_ref_p.memory_size();
2542   int v_align  = vw / elt_size;
2543   assert(v_align > 1, "sanity");
2544   int offset   = align_to_ref_p.offset_in_bytes() / elt_size;
2545   Node *offsn  = _igvn.intcon(offset);
2546 
2547   Node *e = offsn;
2548   if (align_to_ref_p.invar() != NULL) {
2549     // incorporate any extra invariant piece producing (offset +/- invar) >>> log2(elt)
2550     Node* log2_elt = _igvn.intcon(exact_log2(elt_size));
2551     Node* aref     = new URShiftINode(align_to_ref_p.invar(), log2_elt);
2552     _igvn.register_new_node_with_optimizer(aref);
2553     _phase->set_ctrl(aref, pre_ctrl);
2554     if (align_to_ref_p.negate_invar()) {
2555       e = new SubINode(e, aref);
2556     } else {
2557       e = new AddINode(e, aref);
2558     }
2559     _igvn.register_new_node_with_optimizer(e);
2560     _phase->set_ctrl(e, pre_ctrl);
2561   }
2562   if (vw > ObjectAlignmentInBytes) {
2563     // incorporate base e +/- base && Mask >>> log2(elt)
2564     Node* xbase = new CastP2XNode(NULL, align_to_ref_p.base());
2565     _igvn.register_new_node_with_optimizer(xbase);
2566 #ifdef _LP64
2567     xbase  = new ConvL2INode(xbase);
2568     _igvn.register_new_node_with_optimizer(xbase);
2569 #endif
2570     Node* mask = _igvn.intcon(vw-1);
2571     Node* masked_xbase  = new AndINode(xbase, mask);
2572     _igvn.register_new_node_with_optimizer(masked_xbase);
2573     Node* log2_elt = _igvn.intcon(exact_log2(elt_size));
2574     Node* bref     = new URShiftINode(masked_xbase, log2_elt);
2575     _igvn.register_new_node_with_optimizer(bref);
2576     _phase->set_ctrl(bref, pre_ctrl);
2577     e = new AddINode(e, bref);
2578     _igvn.register_new_node_with_optimizer(e);
2579     _phase->set_ctrl(e, pre_ctrl);
2580   }
2581 
2582   // compute e +/- lim0
2583   if (scale < 0) {
2584     e = new SubINode(e, lim0);
2585   } else {
2586     e = new AddINode(e, lim0);
2587   }
2588   _igvn.register_new_node_with_optimizer(e);
2589   _phase->set_ctrl(e, pre_ctrl);
2590 
2591   if (stride * scale > 0) {
2592     // compute V - (e +/- lim0)
2593     Node* va  = _igvn.intcon(v_align);
2594     e = new SubINode(va, e);
2595     _igvn.register_new_node_with_optimizer(e);
2596     _phase->set_ctrl(e, pre_ctrl);
2597   }
2598   // compute N = (exp) % V
2599   Node* va_msk = _igvn.intcon(v_align - 1);
2600   Node* N = new AndINode(e, va_msk);
2601   _igvn.register_new_node_with_optimizer(N);
2602   _phase->set_ctrl(N, pre_ctrl);
2603 
2604   //   substitute back into (1), so that new limit
2605   //     lim = lim0 + N
2606   Node* lim;
2607   if (stride < 0) {
2608     lim = new SubINode(lim0, N);
2609   } else {
2610     lim = new AddINode(lim0, N);
2611   }
2612   _igvn.register_new_node_with_optimizer(lim);
2613   _phase->set_ctrl(lim, pre_ctrl);
2614   Node* constrained =
2615     (stride > 0) ? (Node*) new MinINode(lim, orig_limit)
2616                  : (Node*) new MaxINode(lim, orig_limit);
2617   _igvn.register_new_node_with_optimizer(constrained);
2618   _phase->set_ctrl(constrained, pre_ctrl);
2619   _igvn.hash_delete(pre_opaq);
2620   pre_opaq->set_req(1, constrained);
2621 }
2622 
2623 //----------------------------get_pre_loop_end---------------------------
2624 // Find pre loop end from main loop.  Returns null if none.
2625 CountedLoopEndNode* SuperWord::get_pre_loop_end(CountedLoopNode *cl) {
2626   Node *ctrl = cl->in(LoopNode::EntryControl);
2627   if (!ctrl->is_IfTrue() && !ctrl->is_IfFalse()) return NULL;
2628   Node *iffm = ctrl->in(0);
2629   if (!iffm->is_If()) return NULL;
2630   Node *p_f = iffm->in(0);
2631   if (!p_f->is_IfFalse()) return NULL;
2632   if (!p_f->in(0)->is_CountedLoopEnd()) return NULL;
2633   CountedLoopEndNode *pre_end = p_f->in(0)->as_CountedLoopEnd();
2634   CountedLoopNode* loop_node = pre_end->loopnode();
2635   if (loop_node == NULL || !loop_node->is_pre_loop()) return NULL;
2636   return pre_end;
2637 }
2638 
2639 
2640 //------------------------------init---------------------------
2641 void SuperWord::init() {
2642   _dg.init();
2643   _packset.clear();
2644   _disjoint_ptrs.clear();
2645   _block.clear();
2646   _data_entry.clear();
2647   _mem_slice_head.clear();
2648   _mem_slice_tail.clear();
2649   _iteration_first.clear();
2650   _iteration_last.clear();
2651   _node_info.clear();
2652   _align_to_ref = NULL;
2653   _lpt = NULL;
2654   _lp = NULL;
2655   _bb = NULL;
2656   _iv = NULL;
2657   _race_possible = 0;
2658   _early_return = false;
2659   _num_work_vecs = 0;
2660   _num_reductions = 0;
2661 }
2662 
2663 //------------------------------restart---------------------------
2664 void SuperWord::restart() {
2665   _dg.init();
2666   _packset.clear();
2667   _disjoint_ptrs.clear();
2668   _block.clear();
2669   _data_entry.clear();
2670   _mem_slice_head.clear();
2671   _mem_slice_tail.clear();
2672   _node_info.clear();
2673 }
2674 
2675 //------------------------------print_packset---------------------------
2676 void SuperWord::print_packset() {
2677 #ifndef PRODUCT
2678   tty->print_cr("packset");
2679   for (int i = 0; i < _packset.length(); i++) {
2680     tty->print_cr("Pack: %d", i);
2681     Node_List* p = _packset.at(i);
2682     print_pack(p);
2683   }
2684 #endif
2685 }
2686 
2687 //------------------------------print_pack---------------------------
2688 void SuperWord::print_pack(Node_List* p) {
2689   for (uint i = 0; i < p->size(); i++) {
2690     print_stmt(p->at(i));
2691   }
2692 }
2693 
2694 //------------------------------print_bb---------------------------
2695 void SuperWord::print_bb() {
2696 #ifndef PRODUCT
2697   tty->print_cr("\nBlock");
2698   for (int i = 0; i < _block.length(); i++) {
2699     Node* n = _block.at(i);
2700     tty->print("%d ", i);
2701     if (n) {
2702       n->dump();
2703     }
2704   }
2705 #endif
2706 }
2707 
2708 //------------------------------print_stmt---------------------------
2709 void SuperWord::print_stmt(Node* s) {
2710 #ifndef PRODUCT
2711   tty->print(" align: %d \t", alignment(s));
2712   s->dump();
2713 #endif
2714 }
2715 
2716 //------------------------------blank---------------------------
2717 char* SuperWord::blank(uint depth) {
2718   static char blanks[101];
2719   assert(depth < 101, "too deep");
2720   for (uint i = 0; i < depth; i++) blanks[i] = ' ';
2721   blanks[depth] = '\0';
2722   return blanks;
2723 }
2724 
2725 
2726 //==============================SWPointer===========================
2727 
2728 //----------------------------SWPointer------------------------
2729 SWPointer::SWPointer(MemNode* mem, SuperWord* slp, Node_Stack *nstack, bool analyze_only) :
2730   _mem(mem), _slp(slp),  _base(NULL),  _adr(NULL),
2731   _scale(0), _offset(0), _invar(NULL), _negate_invar(false),
2732   _nstack(nstack), _analyze_only(analyze_only),
2733   _stack_idx(0) {
2734 
2735   Node* adr = mem->in(MemNode::Address);
2736   if (!adr->is_AddP()) {
2737     assert(!valid(), "too complex");
2738     return;
2739   }
2740   // Match AddP(base, AddP(ptr, k*iv [+ invariant]), constant)
2741   Node* base = adr->in(AddPNode::Base);
2742   // The base address should be loop invariant
2743   if (!invariant(base)) {
2744     assert(!valid(), "base address is loop variant");
2745     return;
2746   }
2747   //unsafe reference could not be aligned appropriately without runtime checking
2748   if (base == NULL || base->bottom_type() == Type::TOP) {
2749     assert(!valid(), "unsafe access");
2750     return;
2751   }
2752   for (int i = 0; i < 3; i++) {
2753     if (!scaled_iv_plus_offset(adr->in(AddPNode::Offset))) {
2754       assert(!valid(), "too complex");
2755       return;
2756     }
2757     adr = adr->in(AddPNode::Address);
2758     if (base == adr || !adr->is_AddP()) {
2759       break; // stop looking at addp's
2760     }
2761   }
2762   _base = base;
2763   _adr  = adr;
2764   assert(valid(), "Usable");
2765 }
2766 
2767 // Following is used to create a temporary object during
2768 // the pattern match of an address expression.
2769 SWPointer::SWPointer(SWPointer* p) :
2770   _mem(p->_mem), _slp(p->_slp),  _base(NULL),  _adr(NULL),
2771   _scale(0), _offset(0), _invar(NULL), _negate_invar(false),
2772   _nstack(p->_nstack), _analyze_only(p->_analyze_only),
2773   _stack_idx(p->_stack_idx) {}
2774 
2775 //------------------------scaled_iv_plus_offset--------------------
2776 // Match: k*iv + offset
2777 // where: k is a constant that maybe zero, and
2778 //        offset is (k2 [+/- invariant]) where k2 maybe zero and invariant is optional
2779 bool SWPointer::scaled_iv_plus_offset(Node* n) {
2780   if (scaled_iv(n)) {
2781     return true;
2782   }
2783   if (offset_plus_k(n)) {
2784     return true;
2785   }
2786   int opc = n->Opcode();
2787   if (opc == Op_AddI) {
2788     if (scaled_iv(n->in(1)) && offset_plus_k(n->in(2))) {
2789       return true;
2790     }
2791     if (scaled_iv(n->in(2)) && offset_plus_k(n->in(1))) {
2792       return true;
2793     }
2794   } else if (opc == Op_SubI) {
2795     if (scaled_iv(n->in(1)) && offset_plus_k(n->in(2), true)) {
2796       return true;
2797     }
2798     if (scaled_iv(n->in(2)) && offset_plus_k(n->in(1))) {
2799       _scale *= -1;
2800       return true;
2801     }
2802   }
2803   return false;
2804 }
2805 
2806 //----------------------------scaled_iv------------------------
2807 // Match: k*iv where k is a constant that's not zero
2808 bool SWPointer::scaled_iv(Node* n) {
2809   if (_scale != 0) {
2810     return false;  // already found a scale
2811   }
2812   if (n == iv()) {
2813     _scale = 1;
2814     return true;
2815   }
2816   if (_analyze_only && (invariant(n) == false)) {
2817     _nstack->push(n, _stack_idx++);
2818   }
2819   int opc = n->Opcode();
2820   if (opc == Op_MulI) {
2821     if (n->in(1) == iv() && n->in(2)->is_Con()) {
2822       _scale = n->in(2)->get_int();
2823       return true;
2824     } else if (n->in(2) == iv() && n->in(1)->is_Con()) {
2825       _scale = n->in(1)->get_int();
2826       return true;
2827     }
2828   } else if (opc == Op_LShiftI) {
2829     if (n->in(1) == iv() && n->in(2)->is_Con()) {
2830       _scale = 1 << n->in(2)->get_int();
2831       return true;
2832     }
2833   } else if (opc == Op_ConvI2L) {
2834     if (scaled_iv_plus_offset(n->in(1))) {
2835       return true;
2836     }
2837   } else if (opc == Op_LShiftL) {
2838     if (!has_iv() && _invar == NULL) {
2839       // Need to preserve the current _offset value, so
2840       // create a temporary object for this expression subtree.
2841       // Hacky, so should re-engineer the address pattern match.
2842       SWPointer tmp(this);
2843       if (tmp.scaled_iv_plus_offset(n->in(1))) {
2844         if (tmp._invar == NULL) {
2845           int mult = 1 << n->in(2)->get_int();
2846           _scale   = tmp._scale  * mult;
2847           _offset += tmp._offset * mult;
2848           return true;
2849         }
2850       }
2851     }
2852   }
2853   return false;
2854 }
2855 
2856 //----------------------------offset_plus_k------------------------
2857 // Match: offset is (k [+/- invariant])
2858 // where k maybe zero and invariant is optional, but not both.
2859 bool SWPointer::offset_plus_k(Node* n, bool negate) {
2860   int opc = n->Opcode();
2861   if (opc == Op_ConI) {
2862     _offset += negate ? -(n->get_int()) : n->get_int();
2863     return true;
2864   } else if (opc == Op_ConL) {
2865     // Okay if value fits into an int
2866     const TypeLong* t = n->find_long_type();
2867     if (t->higher_equal(TypeLong::INT)) {
2868       jlong loff = n->get_long();
2869       jint  off  = (jint)loff;
2870       _offset += negate ? -off : loff;
2871       return true;
2872     }
2873     return false;
2874   }
2875   if (_invar != NULL) return false; // already have an invariant
2876   if (_analyze_only && (invariant(n) == false)) {
2877     _nstack->push(n, _stack_idx++);
2878   }
2879   if (opc == Op_AddI) {
2880     if (n->in(2)->is_Con() && invariant(n->in(1))) {
2881       _negate_invar = negate;
2882       _invar = n->in(1);
2883       _offset += negate ? -(n->in(2)->get_int()) : n->in(2)->get_int();
2884       return true;
2885     } else if (n->in(1)->is_Con() && invariant(n->in(2))) {
2886       _offset += negate ? -(n->in(1)->get_int()) : n->in(1)->get_int();
2887       _negate_invar = negate;
2888       _invar = n->in(2);
2889       return true;
2890     }
2891   }
2892   if (opc == Op_SubI) {
2893     if (n->in(2)->is_Con() && invariant(n->in(1))) {
2894       _negate_invar = negate;
2895       _invar = n->in(1);
2896       _offset += !negate ? -(n->in(2)->get_int()) : n->in(2)->get_int();
2897       return true;
2898     } else if (n->in(1)->is_Con() && invariant(n->in(2))) {
2899       _offset += negate ? -(n->in(1)->get_int()) : n->in(1)->get_int();
2900       _negate_invar = !negate;
2901       _invar = n->in(2);
2902       return true;
2903     }
2904   }
2905   if (invariant(n)) {
2906     _negate_invar = negate;
2907     _invar = n;
2908     return true;
2909   }
2910   return false;
2911 }
2912 
2913 //----------------------------print------------------------
2914 void SWPointer::print() {
2915 #ifndef PRODUCT
2916   tty->print("base: %d  adr: %d  scale: %d  offset: %d  invar: %c%d\n",
2917              _base != NULL ? _base->_idx : 0,
2918              _adr  != NULL ? _adr->_idx  : 0,
2919              _scale, _offset,
2920              _negate_invar?'-':'+',
2921              _invar != NULL ? _invar->_idx : 0);
2922 #endif
2923 }
2924 
2925 // ========================= OrderedPair =====================
2926 
2927 const OrderedPair OrderedPair::initial;
2928 
2929 // ========================= SWNodeInfo =====================
2930 
2931 const SWNodeInfo SWNodeInfo::initial;
2932 
2933 
2934 // ============================ DepGraph ===========================
2935 
2936 //------------------------------make_node---------------------------
2937 // Make a new dependence graph node for an ideal node.
2938 DepMem* DepGraph::make_node(Node* node) {
2939   DepMem* m = new (_arena) DepMem(node);
2940   if (node != NULL) {
2941     assert(_map.at_grow(node->_idx) == NULL, "one init only");
2942     _map.at_put_grow(node->_idx, m);
2943   }
2944   return m;
2945 }
2946 
2947 //------------------------------make_edge---------------------------
2948 // Make a new dependence graph edge from dpred -> dsucc
2949 DepEdge* DepGraph::make_edge(DepMem* dpred, DepMem* dsucc) {
2950   DepEdge* e = new (_arena) DepEdge(dpred, dsucc, dsucc->in_head(), dpred->out_head());
2951   dpred->set_out_head(e);
2952   dsucc->set_in_head(e);
2953   return e;
2954 }
2955 
2956 // ========================== DepMem ========================
2957 
2958 //------------------------------in_cnt---------------------------
2959 int DepMem::in_cnt() {
2960   int ct = 0;
2961   for (DepEdge* e = _in_head; e != NULL; e = e->next_in()) ct++;
2962   return ct;
2963 }
2964 
2965 //------------------------------out_cnt---------------------------
2966 int DepMem::out_cnt() {
2967   int ct = 0;
2968   for (DepEdge* e = _out_head; e != NULL; e = e->next_out()) ct++;
2969   return ct;
2970 }
2971 
2972 //------------------------------print-----------------------------
2973 void DepMem::print() {
2974 #ifndef PRODUCT
2975   tty->print("  DepNode %d (", _node->_idx);
2976   for (DepEdge* p = _in_head; p != NULL; p = p->next_in()) {
2977     Node* pred = p->pred()->node();
2978     tty->print(" %d", pred != NULL ? pred->_idx : 0);
2979   }
2980   tty->print(") [");
2981   for (DepEdge* s = _out_head; s != NULL; s = s->next_out()) {
2982     Node* succ = s->succ()->node();
2983     tty->print(" %d", succ != NULL ? succ->_idx : 0);
2984   }
2985   tty->print_cr(" ]");
2986 #endif
2987 }
2988 
2989 // =========================== DepEdge =========================
2990 
2991 //------------------------------DepPreds---------------------------
2992 void DepEdge::print() {
2993 #ifndef PRODUCT
2994   tty->print_cr("DepEdge: %d [ %d ]", _pred->node()->_idx, _succ->node()->_idx);
2995 #endif
2996 }
2997 
2998 // =========================== DepPreds =========================
2999 // Iterator over predecessor edges in the dependence graph.
3000 
3001 //------------------------------DepPreds---------------------------
3002 DepPreds::DepPreds(Node* n, DepGraph& dg) {
3003   _n = n;
3004   _done = false;
3005   if (_n->is_Store() || _n->is_Load()) {
3006     _next_idx = MemNode::Address;
3007     _end_idx  = n->req();
3008     _dep_next = dg.dep(_n)->in_head();
3009   } else if (_n->is_Mem()) {
3010     _next_idx = 0;
3011     _end_idx  = 0;
3012     _dep_next = dg.dep(_n)->in_head();
3013   } else {
3014     _next_idx = 1;
3015     _end_idx  = _n->req();
3016     _dep_next = NULL;
3017   }
3018   next();
3019 }
3020 
3021 //------------------------------next---------------------------
3022 void DepPreds::next() {
3023   if (_dep_next != NULL) {
3024     _current  = _dep_next->pred()->node();
3025     _dep_next = _dep_next->next_in();
3026   } else if (_next_idx < _end_idx) {
3027     _current  = _n->in(_next_idx++);
3028   } else {
3029     _done = true;
3030   }
3031 }
3032 
3033 // =========================== DepSuccs =========================
3034 // Iterator over successor edges in the dependence graph.
3035 
3036 //------------------------------DepSuccs---------------------------
3037 DepSuccs::DepSuccs(Node* n, DepGraph& dg) {
3038   _n = n;
3039   _done = false;
3040   if (_n->is_Load()) {
3041     _next_idx = 0;
3042     _end_idx  = _n->outcnt();
3043     _dep_next = dg.dep(_n)->out_head();
3044   } else if (_n->is_Mem() || _n->is_Phi() && _n->bottom_type() == Type::MEMORY) {
3045     _next_idx = 0;
3046     _end_idx  = 0;
3047     _dep_next = dg.dep(_n)->out_head();
3048   } else {
3049     _next_idx = 0;
3050     _end_idx  = _n->outcnt();
3051     _dep_next = NULL;
3052   }
3053   next();
3054 }
3055 
3056 //-------------------------------next---------------------------
3057 void DepSuccs::next() {
3058   if (_dep_next != NULL) {
3059     _current  = _dep_next->succ()->node();
3060     _dep_next = _dep_next->next_out();
3061   } else if (_next_idx < _end_idx) {
3062     _current  = _n->raw_out(_next_idx++);
3063   } else {
3064     _done = true;
3065   }
3066 }
3067 
3068 //
3069 // --------------------------------- vectorization/simd -----------------------------------
3070 //
3071 Node*  SuperWord::find_phi_for_mem_dep(LoadNode* ld) {
3072   assert(in_bb(ld), "must be in block");
3073   if (_clone_map.gen(ld->_idx) == _ii_first) {
3074 #ifndef PRODUCT
3075     if (_vector_loop_debug) {
3076       tty->print_cr("SuperWord::find_phi_for_mem_dep _clone_map.gen(ld->_idx)=%d",
3077                     _clone_map.gen(ld->_idx));
3078     }
3079 #endif
3080     return NULL; //we think that any ld in the first gen being vectorizable
3081   }
3082 
3083   Node* mem = ld->in(MemNode::Memory);
3084   if (mem->outcnt() <= 1) {
3085     // we don't want to remove the only edge from mem node to load
3086 #ifndef PRODUCT
3087     if (_vector_loop_debug) {
3088       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",
3089                     mem->_idx, ld->_idx);
3090       ld->dump();
3091       mem->dump();
3092     }
3093 #endif
3094     return NULL;
3095   }
3096   if (!in_bb(mem) || _clone_map.gen(mem->_idx) == _clone_map.gen(ld->_idx)) {
3097 #ifndef PRODUCT
3098     if (_vector_loop_debug) {
3099       tty->print_cr("SuperWord::find_phi_for_mem_dep _clone_map.gen(mem->_idx)=%d",
3100                     _clone_map.gen(mem->_idx));
3101     }
3102 #endif
3103     return NULL; // does not depend on loop volatile node or depends on the same generation
3104   }
3105 
3106   //otherwise first node should depend on mem-phi
3107   Node* first = first_node(ld);
3108   assert(first->is_Load(), "must be Load");
3109   Node* phi = first->as_Load()->in(MemNode::Memory);
3110   if (!phi->is_Phi() || phi->bottom_type() != Type::MEMORY) {
3111 #ifndef PRODUCT
3112     if (_vector_loop_debug) {
3113       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");
3114       ld->dump();
3115       first->dump();
3116     }
3117 #endif
3118     return NULL;
3119   }
3120 
3121   Node* tail = 0;
3122   for (int m = 0; m < _mem_slice_head.length(); m++) {
3123     if (_mem_slice_head.at(m) == phi) {
3124       tail = _mem_slice_tail.at(m);
3125     }
3126   }
3127   if (tail == 0) { //test that found phi is in the list  _mem_slice_head
3128 #ifndef PRODUCT
3129     if (_vector_loop_debug) {
3130       tty->print_cr("SuperWord::find_phi_for_mem_dep load %d is not vectorizable node, its phi %d is not _mem_slice_head",
3131                     ld->_idx, phi->_idx);
3132       ld->dump();
3133       phi->dump();
3134     }
3135 #endif
3136     return NULL;
3137   }
3138 
3139   // now all conditions are met
3140   return phi;
3141 }
3142 
3143 Node* SuperWord::first_node(Node* nd) {
3144   for (int ii = 0; ii < _iteration_first.length(); ii++) {
3145     Node* nnn = _iteration_first.at(ii);
3146     if (_clone_map.idx(nnn->_idx) == _clone_map.idx(nd->_idx)) {
3147 #ifndef PRODUCT
3148       if (_vector_loop_debug) {
3149         tty->print_cr("SuperWord::first_node: %d is the first iteration node for %d (_clone_map.idx(nnn->_idx) = %d)",
3150                       nnn->_idx, nd->_idx, _clone_map.idx(nnn->_idx));
3151       }
3152 #endif
3153       return nnn;
3154     }
3155   }
3156 
3157 #ifndef PRODUCT
3158   if (_vector_loop_debug) {
3159     tty->print_cr("SuperWord::first_node: did not find first iteration node for %d (_clone_map.idx(nd->_idx)=%d)",
3160                   nd->_idx, _clone_map.idx(nd->_idx));
3161   }
3162 #endif
3163   return 0;
3164 }
3165 
3166 Node* SuperWord::last_node(Node* nd) {
3167   for (int ii = 0; ii < _iteration_last.length(); ii++) {
3168     Node* nnn = _iteration_last.at(ii);
3169     if (_clone_map.idx(nnn->_idx) == _clone_map.idx(nd->_idx)) {
3170 #ifndef PRODUCT
3171       if (_vector_loop_debug) {
3172         tty->print_cr("SuperWord::last_node _clone_map.idx(nnn->_idx)=%d, _clone_map.idx(nd->_idx)=%d",
3173                       _clone_map.idx(nnn->_idx), _clone_map.idx(nd->_idx));
3174       }
3175 #endif
3176       return nnn;
3177     }
3178   }
3179   return 0;
3180 }
3181 
3182 int SuperWord::mark_generations() {
3183   Node *ii_err = 0, *tail_err;
3184   for (int i = 0; i < _mem_slice_head.length(); i++) {
3185     Node* phi  = _mem_slice_head.at(i);
3186     assert(phi->is_Phi(), "must be phi");
3187 
3188     Node* tail = _mem_slice_tail.at(i);
3189     if (_ii_last == -1) {
3190       tail_err = tail;
3191       _ii_last = _clone_map.gen(tail->_idx);
3192     }
3193     else if (_ii_last != _clone_map.gen(tail->_idx)) {
3194 #ifndef PRODUCT
3195       if (TraceSuperWord && Verbose) {
3196         tty->print_cr("SuperWord::mark_generations _ii_last error - found different generations in two tail nodes ");
3197         tail->dump();
3198         tail_err->dump();
3199       }
3200 #endif
3201       return -1;
3202     }
3203 
3204     // find first iteration in the loop
3205     for (DUIterator_Fast imax, i = phi->fast_outs(imax); i < imax; i++) {
3206       Node* ii = phi->fast_out(i);
3207       if (in_bb(ii) && ii->is_Store()) { // we speculate that normally Stores of one and one only generation have deps from mem phi
3208         if (_ii_first == -1) {
3209           ii_err = ii;
3210           _ii_first = _clone_map.gen(ii->_idx);
3211         } else if (_ii_first != _clone_map.gen(ii->_idx)) {
3212 #ifndef PRODUCT
3213           if (TraceSuperWord && Verbose) {
3214             tty->print_cr("SuperWord::mark_generations _ii_first error - found different generations in two nodes ");
3215             ii->dump();
3216             ii_err->dump();
3217           }
3218 #endif
3219           return -1; // this phi has Stores from different generations of unroll and cannot be simd/vectorized
3220         }
3221       }
3222     }//for (DUIterator_Fast imax,
3223   }//for (int i...
3224 
3225   if (_ii_first == -1 || _ii_last == -1) {
3226 #ifndef PRODUCT
3227     if (TraceSuperWord && Verbose) {
3228       tty->print_cr("SuperWord::mark_generations unknown error, something vent wrong");
3229     }
3230 #endif
3231     return -1; // something vent wrong
3232   }
3233   // collect nodes in the first and last generations
3234   assert(_iteration_first.length() == 0, "_iteration_first must be empty");
3235   assert(_iteration_last.length() == 0, "_iteration_last must be empty");
3236   for (int j = 0; j < _block.length(); j++) {
3237     Node* n = _block.at(j);
3238     node_idx_t gen = _clone_map.gen(n->_idx);
3239     if ((signed)gen == _ii_first) {
3240       _iteration_first.push(n);
3241     } else if ((signed)gen == _ii_last) {
3242       _iteration_last.push(n);
3243     }
3244   }
3245 
3246   // building order of iterations
3247   assert(_ii_order.length() == 0, "should be empty");
3248   if (ii_err != 0) {
3249     assert(in_bb(ii_err) && ii_err->is_Store(), "should be Store in bb");
3250     Node* nd = ii_err;
3251     while(_clone_map.gen(nd->_idx) != _ii_last) {
3252       _ii_order.push(_clone_map.gen(nd->_idx));
3253       bool found = false;
3254       for (DUIterator_Fast imax, i = nd->fast_outs(imax); i < imax; i++) {
3255         Node* use = nd->fast_out(i);
3256         if (_clone_map.idx(use->_idx) == _clone_map.idx(nd->_idx) && use->as_Store()->in(MemNode::Memory) == nd) {
3257           found = true;
3258           nd = use;
3259           break;
3260         }
3261       }//for
3262 
3263       if (found == false) {
3264 #ifndef PRODUCT
3265         if (TraceSuperWord && Verbose) {
3266           tty->print_cr("SuperWord::mark_generations: Cannot build order of iterations - no dependent Store for %d", nd->_idx);
3267         }
3268 #endif
3269         _ii_order.clear();
3270         return -1;
3271       }
3272     } //while
3273     _ii_order.push(_clone_map.gen(nd->_idx));
3274   }
3275 
3276 #ifndef PRODUCT
3277   if (_vector_loop_debug) {
3278     tty->print_cr("SuperWord::mark_generations");
3279     tty->print_cr("First generation (%d) nodes:", _ii_first);
3280     for (int ii = 0; ii < _iteration_first.length(); ii++)  _iteration_first.at(ii)->dump();
3281     tty->print_cr("Last generation (%d) nodes:", _ii_last);
3282     for (int ii = 0; ii < _iteration_last.length(); ii++)  _iteration_last.at(ii)->dump();
3283     tty->print_cr(" ");
3284 
3285     tty->print("SuperWord::List of generations: ");
3286     for (int jj = 0; jj < _ii_order.length(); ++jj) {
3287       tty->print("%d:%d ", jj, _ii_order.at(jj));
3288     }
3289     tty->print_cr(" ");
3290   }
3291 #endif
3292 
3293   return _ii_first;
3294 }
3295 
3296 bool SuperWord::fix_commutative_inputs(Node* gold, Node* fix) {
3297   assert(gold->is_Add() && fix->is_Add() || gold->is_Mul() && fix->is_Mul(), "should be only Add or Mul nodes");
3298   assert(_clone_map.idx(gold->_idx) == _clone_map.idx(fix->_idx), "should be clones of the same node");
3299   Node* gin1 = gold->in(1);
3300   Node* gin2 = gold->in(2);
3301   Node* fin1 = fix->in(1);
3302   Node* fin2 = fix->in(2);
3303   bool swapped = false;
3304 
3305   if (in_bb(gin1) && in_bb(gin2) && in_bb(fin1) && in_bb(fin1)) {
3306     if (_clone_map.idx(gin1->_idx) == _clone_map.idx(fin1->_idx) &&
3307         _clone_map.idx(gin2->_idx) == _clone_map.idx(fin2->_idx)) {
3308       return true; // nothing to fix
3309     }
3310     if (_clone_map.idx(gin1->_idx) == _clone_map.idx(fin2->_idx) &&
3311         _clone_map.idx(gin2->_idx) == _clone_map.idx(fin1->_idx)) {
3312       fix->swap_edges(1, 2);
3313       swapped = true;
3314     }
3315   }
3316   // at least one input comes from outside of bb
3317   if (gin1->_idx == fin1->_idx)  {
3318     return true; // nothing to fix
3319   }
3320   if (!swapped && (gin1->_idx == fin2->_idx || gin2->_idx == fin1->_idx))  { //swapping is expensive, check condition first
3321     fix->swap_edges(1, 2);
3322     swapped = true;
3323   }
3324 
3325   if (swapped) {
3326 #ifndef PRODUCT
3327     if (_vector_loop_debug) {
3328       tty->print_cr("SuperWord::fix_commutative_inputs: fixed node %d", fix->_idx);
3329     }
3330 #endif
3331     return true;
3332   }
3333 
3334 #ifndef PRODUCT
3335   if (TraceSuperWord && Verbose) {
3336     tty->print_cr("SuperWord::fix_commutative_inputs: cannot fix node %d", fix->_idx);
3337   }
3338 #endif
3339   return false;
3340 }
3341 
3342 bool SuperWord::pack_parallel() {
3343 #ifndef PRODUCT
3344   if (_vector_loop_debug) {
3345     tty->print_cr("SuperWord::pack_parallel: START");
3346   }
3347 #endif
3348 
3349   _packset.clear();
3350 
3351   for (int ii = 0; ii < _iteration_first.length(); ii++) {
3352     Node* nd = _iteration_first.at(ii);
3353     if (in_bb(nd) && (nd->is_Load() || nd->is_Store() || nd->is_Add() || nd->is_Mul())) {
3354       Node_List* pk = new Node_List();
3355       pk->push(nd);
3356       for (int gen = 1; gen < _ii_order.length(); ++gen) {
3357         for (int kk = 0; kk < _block.length(); kk++) {
3358           Node* clone = _block.at(kk);
3359           if (_clone_map.idx(clone->_idx) == _clone_map.idx(nd->_idx) &&
3360               _clone_map.gen(clone->_idx) == _ii_order.at(gen)) {
3361             if (nd->is_Add() || nd->is_Mul()) {
3362               fix_commutative_inputs(nd, clone);
3363             }
3364             pk->push(clone);
3365             if (pk->size() == 4) {
3366               _packset.append(pk);
3367 #ifndef PRODUCT
3368               if (_vector_loop_debug) {
3369                 tty->print_cr("SuperWord::pack_parallel: added pack ");
3370                 pk->dump();
3371               }
3372 #endif
3373               if (_clone_map.gen(clone->_idx) != _ii_last) {
3374                 pk = new Node_List();
3375               }
3376             }
3377             break;
3378           }
3379         }
3380       }//for
3381     }//if
3382   }//for
3383 
3384 #ifndef PRODUCT
3385   if (_vector_loop_debug) {
3386     tty->print_cr("SuperWord::pack_parallel: END");
3387   }
3388 #endif
3389 
3390   return true;
3391 }
3392 
3393 bool SuperWord::hoist_loads_in_graph() {
3394   GrowableArray<Node*> loads;
3395 
3396 #ifndef PRODUCT
3397   if (_vector_loop_debug) {
3398     tty->print_cr("SuperWord::hoist_loads_in_graph: total number _mem_slice_head.length() = %d", _mem_slice_head.length());
3399   }
3400 #endif
3401 
3402   for (int i = 0; i < _mem_slice_head.length(); i++) {
3403     Node* n = _mem_slice_head.at(i);
3404     if ( !in_bb(n) || !n->is_Phi() || n->bottom_type() != Type::MEMORY) {
3405 #ifndef PRODUCT
3406       if (TraceSuperWord && Verbose) {
3407         tty->print_cr("SuperWord::hoist_loads_in_graph: skipping unexpected node n=%d", n->_idx);
3408       }
3409 #endif
3410       continue;
3411     }
3412 
3413 #ifndef PRODUCT
3414     if (_vector_loop_debug) {
3415       tty->print_cr("SuperWord::hoist_loads_in_graph: processing phi %d  = _mem_slice_head.at(%d);", n->_idx, i);
3416     }
3417 #endif
3418 
3419     for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3420       Node* ld = n->fast_out(i);
3421       if (ld->is_Load() && ld->as_Load()->in(MemNode::Memory) == n && in_bb(ld)) {
3422         for (int i = 0; i < _block.length(); i++) {
3423           Node* ld2 = _block.at(i);
3424           if (ld2->is_Load() &&
3425               _clone_map.idx(ld->_idx) == _clone_map.idx(ld2->_idx) &&
3426               _clone_map.gen(ld->_idx) != _clone_map.gen(ld2->_idx)) { // <= do not collect the first generation ld
3427 #ifndef PRODUCT
3428             if (_vector_loop_debug) {
3429               tty->print_cr("SuperWord::hoist_loads_in_graph: will try to hoist load ld2->_idx=%d, cloned from %d (ld->_idx=%d)",
3430                             ld2->_idx, _clone_map.idx(ld->_idx), ld->_idx);
3431             }
3432 #endif
3433             // could not do on-the-fly, since iterator is immutable
3434             loads.push(ld2);
3435           }
3436         }// for
3437       }//if
3438     }//for (DUIterator_Fast imax,
3439   }//for (int i = 0; i
3440 
3441   for (int i = 0; i < loads.length(); i++) {
3442     LoadNode* ld = loads.at(i)->as_Load();
3443     Node* phi = find_phi_for_mem_dep(ld);
3444     if (phi != NULL) {
3445 #ifndef PRODUCT
3446       if (_vector_loop_debug) {
3447         tty->print_cr("SuperWord::hoist_loads_in_graph replacing MemNode::Memory(%d) edge in %d with one from %d",
3448                       MemNode::Memory, ld->_idx, phi->_idx);
3449       }
3450 #endif
3451       _igvn.replace_input_of(ld, MemNode::Memory, phi);
3452     }
3453   }//for
3454 
3455   restart(); // invalidate all basic structures, since we rebuilt the graph
3456 
3457 #ifndef PRODUCT
3458   if (TraceSuperWord && Verbose) {
3459     tty->print_cr("\nSuperWord::hoist_loads_in_graph() the graph was rebuilt, all structures invalidated and need rebuild");
3460   }
3461 #endif
3462   return true;
3463 }
3464