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 #ifndef SHARE_VM_OPTO_SUPERWORD_HPP 25 #define SHARE_VM_OPTO_SUPERWORD_HPP 26 27 #include "opto/loopnode.hpp" 28 #include "opto/node.hpp" 29 #include "opto/phaseX.hpp" 30 #include "opto/vectornode.hpp" 31 #include "utilities/growableArray.hpp" 32 33 // 34 // S U P E R W O R D T R A N S F O R M 35 // 36 // SuperWords are short, fixed length vectors. 37 // 38 // Algorithm from: 39 // 40 // Exploiting SuperWord Level Parallelism with 41 // Multimedia Instruction Sets 42 // by 43 // Samuel Larsen and Saman Amarasinghe 44 // MIT Laboratory for Computer Science 45 // date 46 // May 2000 47 // published in 48 // ACM SIGPLAN Notices 49 // Proceedings of ACM PLDI '00, Volume 35 Issue 5 50 // 51 // Definition 3.1 A Pack is an n-tuple, <s1, ...,sn>, where 52 // s1,...,sn are independent isomorphic statements in a basic 53 // block. 54 // 55 // Definition 3.2 A PackSet is a set of Packs. 56 // 57 // Definition 3.3 A Pair is a Pack of size two, where the 58 // first statement is considered the left element, and the 59 // second statement is considered the right element. 60 61 class SWPointer; 62 class OrderedPair; 63 64 // ========================= Dependence Graph ===================== 65 66 class DepMem; 67 68 //------------------------------DepEdge--------------------------- 69 // An edge in the dependence graph. The edges incident to a dependence 70 // node are threaded through _next_in for incoming edges and _next_out 71 // for outgoing edges. 72 class DepEdge : public ResourceObj { 73 protected: 74 DepMem* _pred; 75 DepMem* _succ; 76 DepEdge* _next_in; // list of in edges, null terminated 77 DepEdge* _next_out; // list of out edges, null terminated 78 79 public: 80 DepEdge(DepMem* pred, DepMem* succ, DepEdge* next_in, DepEdge* next_out) : 81 _pred(pred), _succ(succ), _next_in(next_in), _next_out(next_out) {} 82 83 DepEdge* next_in() { return _next_in; } 84 DepEdge* next_out() { return _next_out; } 85 DepMem* pred() { return _pred; } 86 DepMem* succ() { return _succ; } 87 88 void print(); 89 }; 90 91 //------------------------------DepMem--------------------------- 92 // A node in the dependence graph. _in_head starts the threaded list of 93 // incoming edges, and _out_head starts the list of outgoing edges. 94 class DepMem : public ResourceObj { 95 protected: 96 Node* _node; // Corresponding ideal node 97 DepEdge* _in_head; // Head of list of in edges, null terminated 98 DepEdge* _out_head; // Head of list of out edges, null terminated 99 100 public: 101 DepMem(Node* node) : _node(node), _in_head(NULL), _out_head(NULL) {} 102 103 Node* node() { return _node; } 104 DepEdge* in_head() { return _in_head; } 105 DepEdge* out_head() { return _out_head; } 106 void set_in_head(DepEdge* hd) { _in_head = hd; } 107 void set_out_head(DepEdge* hd) { _out_head = hd; } 108 109 int in_cnt(); // Incoming edge count 110 int out_cnt(); // Outgoing edge count 111 112 void print(); 113 }; 114 115 //------------------------------DepGraph--------------------------- 116 class DepGraph VALUE_OBJ_CLASS_SPEC { 117 protected: 118 Arena* _arena; 119 GrowableArray<DepMem*> _map; 120 DepMem* _root; 121 DepMem* _tail; 122 123 public: 124 DepGraph(Arena* a) : _arena(a), _map(a, 8, 0, NULL) { 125 _root = new (_arena) DepMem(NULL); 126 _tail = new (_arena) DepMem(NULL); 127 } 128 129 DepMem* root() { return _root; } 130 DepMem* tail() { return _tail; } 131 132 // Return dependence node corresponding to an ideal node 133 DepMem* dep(Node* node) { return _map.at(node->_idx); } 134 135 // Make a new dependence graph node for an ideal node. 136 DepMem* make_node(Node* node); 137 138 // Make a new dependence graph edge dprec->dsucc 139 DepEdge* make_edge(DepMem* dpred, DepMem* dsucc); 140 141 DepEdge* make_edge(Node* pred, Node* succ) { return make_edge(dep(pred), dep(succ)); } 142 DepEdge* make_edge(DepMem* pred, Node* succ) { return make_edge(pred, dep(succ)); } 143 DepEdge* make_edge(Node* pred, DepMem* succ) { return make_edge(dep(pred), succ); } 144 145 void init() { _map.clear(); } // initialize 146 147 void print(Node* n) { dep(n)->print(); } 148 void print(DepMem* d) { d->print(); } 149 }; 150 151 //------------------------------DepPreds--------------------------- 152 // Iterator over predecessors in the dependence graph and 153 // non-memory-graph inputs of ideal nodes. 154 class DepPreds : public StackObj { 155 private: 156 Node* _n; 157 int _next_idx, _end_idx; 158 DepEdge* _dep_next; 159 Node* _current; 160 bool _done; 161 162 public: 163 DepPreds(Node* n, DepGraph& dg); 164 Node* current() { return _current; } 165 bool done() { return _done; } 166 void next(); 167 }; 168 169 //------------------------------DepSuccs--------------------------- 170 // Iterator over successors in the dependence graph and 171 // non-memory-graph outputs of ideal nodes. 172 class DepSuccs : public StackObj { 173 private: 174 Node* _n; 175 int _next_idx, _end_idx; 176 DepEdge* _dep_next; 177 Node* _current; 178 bool _done; 179 180 public: 181 DepSuccs(Node* n, DepGraph& dg); 182 Node* current() { return _current; } 183 bool done() { return _done; } 184 void next(); 185 }; 186 187 188 // ========================= SuperWord ===================== 189 190 // -----------------------------SWNodeInfo--------------------------------- 191 // Per node info needed by SuperWord 192 class SWNodeInfo VALUE_OBJ_CLASS_SPEC { 193 public: 194 int _alignment; // memory alignment for a node 195 int _depth; // Max expression (DAG) depth from block start 196 const Type* _velt_type; // vector element type 197 Node_List* _my_pack; // pack containing this node 198 199 SWNodeInfo() : _alignment(-1), _depth(0), _velt_type(NULL), _my_pack(NULL) {} 200 static const SWNodeInfo initial; 201 }; 202 203 // -----------------------------SuperWord--------------------------------- 204 // Transforms scalar operations into packed (superword) operations. 205 class SuperWord : public ResourceObj { 206 private: 207 PhaseIdealLoop* _phase; 208 Arena* _arena; 209 PhaseIterGVN &_igvn; 210 211 enum consts { top_align = -1, bottom_align = -666 }; 212 213 GrowableArray<Node_List*> _packset; // Packs for the current block 214 215 GrowableArray<int> _bb_idx; // Map from Node _idx to index within block 216 217 GrowableArray<Node*> _block; // Nodes in current block 218 GrowableArray<Node*> _data_entry; // Nodes with all inputs from outside 219 GrowableArray<Node*> _mem_slice_head; // Memory slice head nodes 220 GrowableArray<Node*> _mem_slice_tail; // Memory slice tail nodes 221 GrowableArray<Node*> _iteration_first; // nodes in the generation that has deps from phi 222 GrowableArray<Node*> _iteration_last; // nodes in the generation that has deps to phi 223 GrowableArray<SWNodeInfo> _node_info; // Info needed per node 224 CloneMap& _clone_map; // map of nodes created in cloning 225 226 MemNode* _align_to_ref; // Memory reference that pre-loop will align to 227 228 GrowableArray<OrderedPair> _disjoint_ptrs; // runtime disambiguated pointer pairs 229 230 DepGraph _dg; // Dependence graph 231 232 // Scratch pads 233 VectorSet _visited; // Visited set 234 VectorSet _post_visited; // Post-visited set 235 Node_Stack _n_idx_list; // List of (node,index) pairs 236 GrowableArray<Node*> _nlist; // List of nodes 237 GrowableArray<Node*> _stk; // Stack of nodes 238 239 public: 240 SuperWord(PhaseIdealLoop* phase); 241 242 void transform_loop(IdealLoopTree* lpt, bool do_optimization); 243 244 void unrolling_analysis(CountedLoopNode *cl, int &local_loop_unroll_factor); 245 246 // Accessors for SWPointer 247 PhaseIdealLoop* phase() { return _phase; } 248 IdealLoopTree* lpt() { return _lpt; } 249 PhiNode* iv() { return _iv; } 250 bool early_return() { return _early_return; } 251 252 private: 253 IdealLoopTree* _lpt; // Current loop tree node 254 LoopNode* _lp; // Current LoopNode 255 Node* _bb; // Current basic block 256 PhiNode* _iv; // Induction var 257 bool _race_possible; // In cases where SDMU is true 258 bool _early_return; // True if we do not initialize 259 bool _do_vector_loop; // whether to do vectorization/simd style 260 bool _vector_loop_debug; // provide more printing in debug mode 261 int _num_work_vecs; // Number of non memory vector operations 262 int _num_reductions; // Number of reduction expressions applied 263 int _ii_first; // generation with direct deps from mem phi 264 int _ii_last; // generation with direct deps to mem phi 265 GrowableArray<int> _ii_order; 266 267 // Accessors 268 Arena* arena() { return _arena; } 269 270 Node* bb() { return _bb; } 271 void set_bb(Node* bb) { _bb = bb; } 272 273 void set_lpt(IdealLoopTree* lpt) { _lpt = lpt; } 274 275 LoopNode* lp() { return _lp; } 276 void set_lp(LoopNode* lp) { _lp = lp; 277 _iv = lp->as_CountedLoop()->phi()->as_Phi(); } 278 int iv_stride() { return lp()->as_CountedLoop()->stride_con(); } 279 280 int vector_width(Node* n) { 281 BasicType bt = velt_basic_type(n); 282 return MIN2(ABS(iv_stride()), Matcher::max_vector_size(bt)); 283 } 284 int vector_width_in_bytes(Node* n) { 285 BasicType bt = velt_basic_type(n); 286 return vector_width(n)*type2aelembytes(bt); 287 } 288 MemNode* align_to_ref() { return _align_to_ref; } 289 void set_align_to_ref(MemNode* m) { _align_to_ref = m; } 290 291 Node* ctrl(Node* n) const { return _phase->has_ctrl(n) ? _phase->get_ctrl(n) : n; } 292 293 // block accessors 294 bool in_bb(Node* n) { return n != NULL && n->outcnt() > 0 && ctrl(n) == _bb; } 295 int bb_idx(Node* n) { assert(in_bb(n), "must be"); return _bb_idx.at(n->_idx); } 296 void set_bb_idx(Node* n, int i) { _bb_idx.at_put_grow(n->_idx, i); } 297 298 // visited set accessors 299 void visited_clear() { _visited.Clear(); } 300 void visited_set(Node* n) { return _visited.set(bb_idx(n)); } 301 int visited_test(Node* n) { return _visited.test(bb_idx(n)); } 302 int visited_test_set(Node* n) { return _visited.test_set(bb_idx(n)); } 303 void post_visited_clear() { _post_visited.Clear(); } 304 void post_visited_set(Node* n) { return _post_visited.set(bb_idx(n)); } 305 int post_visited_test(Node* n) { return _post_visited.test(bb_idx(n)); } 306 307 // Ensure node_info contains element "i" 308 void grow_node_info(int i) { if (i >= _node_info.length()) _node_info.at_put_grow(i, SWNodeInfo::initial); } 309 310 // memory alignment for a node 311 int alignment(Node* n) { return _node_info.adr_at(bb_idx(n))->_alignment; } 312 void set_alignment(Node* n, int a) { int i = bb_idx(n); grow_node_info(i); _node_info.adr_at(i)->_alignment = a; } 313 314 // Max expression (DAG) depth from beginning of the block for each node 315 int depth(Node* n) { return _node_info.adr_at(bb_idx(n))->_depth; } 316 void set_depth(Node* n, int d) { int i = bb_idx(n); grow_node_info(i); _node_info.adr_at(i)->_depth = d; } 317 318 // vector element type 319 const Type* velt_type(Node* n) { return _node_info.adr_at(bb_idx(n))->_velt_type; } 320 BasicType velt_basic_type(Node* n) { return velt_type(n)->array_element_basic_type(); } 321 void set_velt_type(Node* n, const Type* t) { int i = bb_idx(n); grow_node_info(i); _node_info.adr_at(i)->_velt_type = t; } 322 bool same_velt_type(Node* n1, Node* n2); 323 324 // my_pack 325 Node_List* my_pack(Node* n) { return !in_bb(n) ? NULL : _node_info.adr_at(bb_idx(n))->_my_pack; } 326 void set_my_pack(Node* n, Node_List* p) { int i = bb_idx(n); grow_node_info(i); _node_info.adr_at(i)->_my_pack = p; } 327 328 // methods 329 330 // Extract the superword level parallelism 331 void SLP_extract(); 332 // Find the adjacent memory references and create pack pairs for them. 333 void find_adjacent_refs(); 334 // Find a memory reference to align the loop induction variable to. 335 MemNode* find_align_to_ref(Node_List &memops); 336 // Calculate loop's iv adjustment for this memory ops. 337 int get_iv_adjustment(MemNode* mem); 338 // Can the preloop align the reference to position zero in the vector? 339 bool ref_is_alignable(SWPointer& p); 340 // rebuild the graph so all loads in different iterations of cloned loop become dependant on phi node (in _do_vector_loop only) 341 bool hoist_loads_in_graph(); 342 // Test whether MemNode::Memory dependency to the same load but in the first iteration of this loop is coming from memory phi 343 // Return false if failed. 344 Node* find_phi_for_mem_dep(LoadNode* ld); 345 // Return same node but from the first generation. Return 0, if not found 346 Node* first_node(Node* nd); 347 // Return same node as this but from the last generation. Return 0, if not found 348 Node* last_node(Node* n); 349 // Mark nodes belonging to first and last generation, 350 // returns first generation index or -1 if vectorization/simd is impossible 351 int mark_generations(); 352 // swapping inputs of commutative instruction (Add or Mul) 353 bool fix_commutative_inputs(Node* gold, Node* fix); 354 // make packs forcefully (in _do_vector_loop only) 355 bool pack_parallel(); 356 // Construct dependency graph. 357 void dependence_graph(); 358 // Return a memory slice (node list) in predecessor order starting at "start" 359 void mem_slice_preds(Node* start, Node* stop, GrowableArray<Node*> &preds); 360 // Can s1 and s2 be in a pack with s1 immediately preceding s2 and s1 aligned at "align" 361 bool stmts_can_pack(Node* s1, Node* s2, int align); 362 // Does s exist in a pack at position pos? 363 bool exists_at(Node* s, uint pos); 364 // Is s1 immediately before s2 in memory? 365 bool are_adjacent_refs(Node* s1, Node* s2); 366 // Are s1 and s2 similar? 367 bool isomorphic(Node* s1, Node* s2); 368 // Is there no data path from s1 to s2 or s2 to s1? 369 bool independent(Node* s1, Node* s2); 370 // Is there a data path between s1 and s2 and both are reductions? 371 bool reduction(Node* s1, Node* s2); 372 // Helper for independent 373 bool independent_path(Node* shallow, Node* deep, uint dp=0); 374 void set_alignment(Node* s1, Node* s2, int align); 375 int data_size(Node* s); 376 // Extend packset by following use->def and def->use links from pack members. 377 void extend_packlist(); 378 // Extend the packset by visiting operand definitions of nodes in pack p 379 bool follow_use_defs(Node_List* p); 380 // Extend the packset by visiting uses of nodes in pack p 381 bool follow_def_uses(Node_List* p); 382 // For extended packsets, ordinally arrange uses packset by major component 383 void order_def_uses(Node_List* p); 384 // Estimate the savings from executing s1 and s2 as a pack 385 int est_savings(Node* s1, Node* s2); 386 int adjacent_profit(Node* s1, Node* s2); 387 int pack_cost(int ct); 388 int unpack_cost(int ct); 389 // Combine packs A and B with A.last == B.first into A.first..,A.last,B.second,..B.last 390 void combine_packs(); 391 // Construct the map from nodes to packs. 392 void construct_my_pack_map(); 393 // Remove packs that are not implemented or not profitable. 394 void filter_packs(); 395 // Adjust the memory graph for the packed operations 396 void schedule(); 397 // Remove "current" from its current position in the memory graph and insert 398 // it after the appropriate insert points (lip or uip); 399 void remove_and_insert(MemNode *current, MemNode *prev, MemNode *lip, Node *uip, Unique_Node_List &schd_before); 400 // Within a store pack, schedule stores together by moving out the sandwiched memory ops according 401 // to dependence info; and within a load pack, move loads down to the last executed load. 402 void co_locate_pack(Node_List* p); 403 // Convert packs into vector node operations 404 void output(); 405 // Create a vector operand for the nodes in pack p for operand: in(opd_idx) 406 Node* vector_opd(Node_List* p, int opd_idx); 407 // Can code be generated for pack p? 408 bool implemented(Node_List* p); 409 // For pack p, are all operands and all uses (with in the block) vector? 410 bool profitable(Node_List* p); 411 // If a use of pack p is not a vector use, then replace the use with an extract operation. 412 void insert_extracts(Node_List* p); 413 // Is use->in(u_idx) a vector use? 414 bool is_vector_use(Node* use, int u_idx); 415 // Construct reverse postorder list of block members 416 bool construct_bb(); 417 // Initialize per node info 418 void initialize_bb(); 419 // Insert n into block after pos 420 void bb_insert_after(Node* n, int pos); 421 // Compute max depth for expressions from beginning of block 422 void compute_max_depth(); 423 // Compute necessary vector element type for expressions 424 void compute_vector_element_type(); 425 // Are s1 and s2 in a pack pair and ordered as s1,s2? 426 bool in_packset(Node* s1, Node* s2); 427 // Is s in pack p? 428 Node_List* in_pack(Node* s, Node_List* p); 429 // Remove the pack at position pos in the packset 430 void remove_pack_at(int pos); 431 // Return the node executed first in pack p. 432 Node* executed_first(Node_List* p); 433 // Return the node executed last in pack p. 434 Node* executed_last(Node_List* p); 435 // Alignment within a vector memory reference 436 int memory_alignment(MemNode* s, int iv_adjust); 437 // (Start, end] half-open range defining which operands are vector 438 void vector_opd_range(Node* n, uint* start, uint* end); 439 // Smallest type containing range of values 440 const Type* container_type(Node* n); 441 // Adjust pre-loop limit so that in main loop, a load/store reference 442 // to align_to_ref will be a position zero in the vector. 443 void align_initial_loop_index(MemNode* align_to_ref); 444 // Find pre loop end from main loop. Returns null if none. 445 CountedLoopEndNode* get_pre_loop_end(CountedLoopNode *cl); 446 // Is the use of d1 in u1 at the same operand position as d2 in u2? 447 bool opnd_positions_match(Node* d1, Node* u1, Node* d2, Node* u2); 448 void init(); 449 // clean up some basic structures - used if the ideal graph was rebuilt 450 void restart(); 451 452 // print methods 453 void print_packset(); 454 void print_pack(Node_List* p); 455 void print_bb(); 456 void print_stmt(Node* s); 457 char* blank(uint depth); 458 459 void packset_sort(int n); 460 }; 461 462 463 464 //------------------------------SWPointer--------------------------- 465 // Information about an address for dependence checking and vector alignment 466 class SWPointer VALUE_OBJ_CLASS_SPEC { 467 protected: 468 MemNode* _mem; // My memory reference node 469 SuperWord* _slp; // SuperWord class 470 471 Node* _base; // NULL if unsafe nonheap reference 472 Node* _adr; // address pointer 473 jint _scale; // multiplier for iv (in bytes), 0 if no loop iv 474 jint _offset; // constant offset (in bytes) 475 Node* _invar; // invariant offset (in bytes), NULL if none 476 bool _negate_invar; // if true then use: (0 - _invar) 477 Node_Stack* _nstack; // stack used to record a swpointer trace of variants 478 bool _analyze_only; // Used in loop unrolling only for swpointer trace 479 uint _stack_idx; // Used in loop unrolling only for swpointer trace 480 481 PhaseIdealLoop* phase() { return _slp->phase(); } 482 IdealLoopTree* lpt() { return _slp->lpt(); } 483 PhiNode* iv() { return _slp->iv(); } // Induction var 484 485 bool invariant(Node* n) { 486 Node *n_c = phase()->get_ctrl(n); 487 return !lpt()->is_member(phase()->get_loop(n_c)); 488 } 489 490 // Match: k*iv + offset 491 bool scaled_iv_plus_offset(Node* n); 492 // Match: k*iv where k is a constant that's not zero 493 bool scaled_iv(Node* n); 494 // Match: offset is (k [+/- invariant]) 495 bool offset_plus_k(Node* n, bool negate = false); 496 497 public: 498 enum CMP { 499 Less = 1, 500 Greater = 2, 501 Equal = 4, 502 NotEqual = (Less | Greater), 503 NotComparable = (Less | Greater | Equal) 504 }; 505 506 SWPointer(MemNode* mem, SuperWord* slp, Node_Stack *nstack, bool analyze_only); 507 // Following is used to create a temporary object during 508 // the pattern match of an address expression. 509 SWPointer(SWPointer* p); 510 511 bool valid() { return _adr != NULL; } 512 bool has_iv() { return _scale != 0; } 513 514 Node* base() { return _base; } 515 Node* adr() { return _adr; } 516 MemNode* mem() { return _mem; } 517 int scale_in_bytes() { return _scale; } 518 Node* invar() { return _invar; } 519 bool negate_invar() { return _negate_invar; } 520 int offset_in_bytes() { return _offset; } 521 int memory_size() { return _mem->memory_size(); } 522 Node_Stack* node_stack() { return _nstack; } 523 524 // Comparable? 525 int cmp(SWPointer& q) { 526 if (valid() && q.valid() && 527 (_adr == q._adr || _base == _adr && q._base == q._adr) && 528 _scale == q._scale && 529 _invar == q._invar && 530 _negate_invar == q._negate_invar) { 531 bool overlap = q._offset < _offset + memory_size() && 532 _offset < q._offset + q.memory_size(); 533 return overlap ? Equal : (_offset < q._offset ? Less : Greater); 534 } else { 535 return NotComparable; 536 } 537 } 538 539 bool not_equal(SWPointer& q) { return not_equal(cmp(q)); } 540 bool equal(SWPointer& q) { return equal(cmp(q)); } 541 bool comparable(SWPointer& q) { return comparable(cmp(q)); } 542 static bool not_equal(int cmp) { return cmp <= NotEqual; } 543 static bool equal(int cmp) { return cmp == Equal; } 544 static bool comparable(int cmp) { return cmp < NotComparable; } 545 546 void print(); 547 }; 548 549 550 //------------------------------OrderedPair--------------------------- 551 // Ordered pair of Node*. 552 class OrderedPair VALUE_OBJ_CLASS_SPEC { 553 protected: 554 Node* _p1; 555 Node* _p2; 556 public: 557 OrderedPair() : _p1(NULL), _p2(NULL) {} 558 OrderedPair(Node* p1, Node* p2) { 559 if (p1->_idx < p2->_idx) { 560 _p1 = p1; _p2 = p2; 561 } else { 562 _p1 = p2; _p2 = p1; 563 } 564 } 565 566 bool operator==(const OrderedPair &rhs) { 567 return _p1 == rhs._p1 && _p2 == rhs._p2; 568 } 569 void print() { tty->print(" (%d, %d)", _p1->_idx, _p2->_idx); } 570 571 static const OrderedPair initial; 572 }; 573 574 #endif // SHARE_VM_OPTO_SUPERWORD_HPP