rev 7350 : 8078497: C2's superword optimization causes unaligned memory accesses Summary: Prevent vectorization of memory operations with different invariant offsets if unaligned memory accesses are not allowed. Reviewed-by: kvn
1 /* 2 * Copyright (c) 2007, 2013, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 */ 23 24 #ifndef SHARE_VM_OPTO_SUPERWORD_HPP 25 #define SHARE_VM_OPTO_SUPERWORD_HPP 26 27 #include "opto/connode.hpp" 28 #include "opto/loopnode.hpp" 29 #include "opto/node.hpp" 30 #include "opto/phaseX.hpp" 31 #include "opto/vectornode.hpp" 32 #include "utilities/growableArray.hpp" 33 34 // 35 // S U P E R W O R D T R A N S F O R M 36 // 37 // SuperWords are short, fixed length vectors. 38 // 39 // Algorithm from: 40 // 41 // Exploiting SuperWord Level Parallelism with 42 // Multimedia Instruction Sets 43 // by 44 // Samuel Larsen and Saman Amarasinghe 45 // MIT Laboratory for Computer Science 46 // date 47 // May 2000 48 // published in 49 // ACM SIGPLAN Notices 50 // Proceedings of ACM PLDI '00, Volume 35 Issue 5 51 // 52 // Definition 3.1 A Pack is an n-tuple, <s1, ...,sn>, where 53 // s1,...,sn are independent isomorphic statements in a basic 54 // block. 55 // 56 // Definition 3.2 A PackSet is a set of Packs. 57 // 58 // Definition 3.3 A Pair is a Pack of size two, where the 59 // first statement is considered the left element, and the 60 // second statement is considered the right element. 61 62 class SWPointer; 63 class OrderedPair; 64 65 // ========================= Dependence Graph ===================== 66 67 class DepMem; 68 69 //------------------------------DepEdge--------------------------- 70 // An edge in the dependence graph. The edges incident to a dependence 71 // node are threaded through _next_in for incoming edges and _next_out 72 // for outgoing edges. 73 class DepEdge : public ResourceObj { 74 protected: 75 DepMem* _pred; 76 DepMem* _succ; 77 DepEdge* _next_in; // list of in edges, null terminated 78 DepEdge* _next_out; // list of out edges, null terminated 79 80 public: 81 DepEdge(DepMem* pred, DepMem* succ, DepEdge* next_in, DepEdge* next_out) : 82 _pred(pred), _succ(succ), _next_in(next_in), _next_out(next_out) {} 83 84 DepEdge* next_in() { return _next_in; } 85 DepEdge* next_out() { return _next_out; } 86 DepMem* pred() { return _pred; } 87 DepMem* succ() { return _succ; } 88 89 void print(); 90 }; 91 92 //------------------------------DepMem--------------------------- 93 // A node in the dependence graph. _in_head starts the threaded list of 94 // incoming edges, and _out_head starts the list of outgoing edges. 95 class DepMem : public ResourceObj { 96 protected: 97 Node* _node; // Corresponding ideal node 98 DepEdge* _in_head; // Head of list of in edges, null terminated 99 DepEdge* _out_head; // Head of list of out edges, null terminated 100 101 public: 102 DepMem(Node* node) : _node(node), _in_head(NULL), _out_head(NULL) {} 103 104 Node* node() { return _node; } 105 DepEdge* in_head() { return _in_head; } 106 DepEdge* out_head() { return _out_head; } 107 void set_in_head(DepEdge* hd) { _in_head = hd; } 108 void set_out_head(DepEdge* hd) { _out_head = hd; } 109 110 int in_cnt(); // Incoming edge count 111 int out_cnt(); // Outgoing edge count 112 113 void print(); 114 }; 115 116 //------------------------------DepGraph--------------------------- 117 class DepGraph VALUE_OBJ_CLASS_SPEC { 118 protected: 119 Arena* _arena; 120 GrowableArray<DepMem*> _map; 121 DepMem* _root; 122 DepMem* _tail; 123 124 public: 125 DepGraph(Arena* a) : _arena(a), _map(a, 8, 0, NULL) { 126 _root = new (_arena) DepMem(NULL); 127 _tail = new (_arena) DepMem(NULL); 128 } 129 130 DepMem* root() { return _root; } 131 DepMem* tail() { return _tail; } 132 133 // Return dependence node corresponding to an ideal node 134 DepMem* dep(Node* node) { return _map.at(node->_idx); } 135 136 // Make a new dependence graph node for an ideal node. 137 DepMem* make_node(Node* node); 138 139 // Make a new dependence graph edge dprec->dsucc 140 DepEdge* make_edge(DepMem* dpred, DepMem* dsucc); 141 142 DepEdge* make_edge(Node* pred, Node* succ) { return make_edge(dep(pred), dep(succ)); } 143 DepEdge* make_edge(DepMem* pred, Node* succ) { return make_edge(pred, dep(succ)); } 144 DepEdge* make_edge(Node* pred, DepMem* succ) { return make_edge(dep(pred), succ); } 145 146 void init() { _map.clear(); } // initialize 147 148 void print(Node* n) { dep(n)->print(); } 149 void print(DepMem* d) { d->print(); } 150 }; 151 152 //------------------------------DepPreds--------------------------- 153 // Iterator over predecessors in the dependence graph and 154 // non-memory-graph inputs of ideal nodes. 155 class DepPreds : public StackObj { 156 private: 157 Node* _n; 158 int _next_idx, _end_idx; 159 DepEdge* _dep_next; 160 Node* _current; 161 bool _done; 162 163 public: 164 DepPreds(Node* n, DepGraph& dg); 165 Node* current() { return _current; } 166 bool done() { return _done; } 167 void next(); 168 }; 169 170 //------------------------------DepSuccs--------------------------- 171 // Iterator over successors in the dependence graph and 172 // non-memory-graph outputs of ideal nodes. 173 class DepSuccs : public StackObj { 174 private: 175 Node* _n; 176 int _next_idx, _end_idx; 177 DepEdge* _dep_next; 178 Node* _current; 179 bool _done; 180 181 public: 182 DepSuccs(Node* n, DepGraph& dg); 183 Node* current() { return _current; } 184 bool done() { return _done; } 185 void next(); 186 }; 187 188 189 // ========================= SuperWord ===================== 190 191 // -----------------------------SWNodeInfo--------------------------------- 192 // Per node info needed by SuperWord 193 class SWNodeInfo VALUE_OBJ_CLASS_SPEC { 194 public: 195 int _alignment; // memory alignment for a node 196 int _depth; // Max expression (DAG) depth from block start 197 const Type* _velt_type; // vector element type 198 Node_List* _my_pack; // pack containing this node 199 200 SWNodeInfo() : _alignment(-1), _depth(0), _velt_type(NULL), _my_pack(NULL) {} 201 static const SWNodeInfo initial; 202 }; 203 204 // -----------------------------SuperWord--------------------------------- 205 // Transforms scalar operations into packed (superword) operations. 206 class SuperWord : public ResourceObj { 207 private: 208 PhaseIdealLoop* _phase; 209 Arena* _arena; 210 PhaseIterGVN &_igvn; 211 212 enum consts { top_align = -1, bottom_align = -666 }; 213 214 GrowableArray<Node_List*> _packset; // Packs for the current block 215 216 GrowableArray<int> _bb_idx; // Map from Node _idx to index within block 217 218 GrowableArray<Node*> _block; // Nodes in current block 219 GrowableArray<Node*> _data_entry; // Nodes with all inputs from outside 220 GrowableArray<Node*> _mem_slice_head; // Memory slice head nodes 221 GrowableArray<Node*> _mem_slice_tail; // Memory slice tail nodes 222 223 GrowableArray<SWNodeInfo> _node_info; // Info needed per node 224 225 MemNode* _align_to_ref; // Memory reference that pre-loop will align to 226 227 GrowableArray<OrderedPair> _disjoint_ptrs; // runtime disambiguated pointer pairs 228 229 DepGraph _dg; // Dependence graph 230 231 // Scratch pads 232 VectorSet _visited; // Visited set 233 VectorSet _post_visited; // Post-visited set 234 Node_Stack _n_idx_list; // List of (node,index) pairs 235 GrowableArray<Node*> _nlist; // List of nodes 236 GrowableArray<Node*> _stk; // Stack of nodes 237 238 public: 239 SuperWord(PhaseIdealLoop* phase); 240 241 void transform_loop(IdealLoopTree* lpt); 242 243 // Accessors for SWPointer 244 PhaseIdealLoop* phase() { return _phase; } 245 IdealLoopTree* lpt() { return _lpt; } 246 PhiNode* iv() { return _iv; } 247 248 private: 249 IdealLoopTree* _lpt; // Current loop tree node 250 LoopNode* _lp; // Current LoopNode 251 Node* _bb; // Current basic block 252 PhiNode* _iv; // Induction var 253 254 // Accessors 255 Arena* arena() { return _arena; } 256 257 Node* bb() { return _bb; } 258 void set_bb(Node* bb) { _bb = bb; } 259 260 void set_lpt(IdealLoopTree* lpt) { _lpt = lpt; } 261 262 LoopNode* lp() { return _lp; } 263 void set_lp(LoopNode* lp) { _lp = lp; 264 _iv = lp->as_CountedLoop()->phi()->as_Phi(); } 265 int iv_stride() { return lp()->as_CountedLoop()->stride_con(); } 266 267 int vector_width(Node* n) { 268 BasicType bt = velt_basic_type(n); 269 return MIN2(ABS(iv_stride()), Matcher::max_vector_size(bt)); 270 } 271 int vector_width_in_bytes(Node* n) { 272 BasicType bt = velt_basic_type(n); 273 return vector_width(n)*type2aelembytes(bt); 274 } 275 MemNode* align_to_ref() { return _align_to_ref; } 276 void set_align_to_ref(MemNode* m) { _align_to_ref = m; } 277 278 Node* ctrl(Node* n) const { return _phase->has_ctrl(n) ? _phase->get_ctrl(n) : n; } 279 280 // block accessors 281 bool in_bb(Node* n) { return n != NULL && n->outcnt() > 0 && ctrl(n) == _bb; } 282 int bb_idx(Node* n) { assert(in_bb(n), "must be"); return _bb_idx.at(n->_idx); } 283 void set_bb_idx(Node* n, int i) { _bb_idx.at_put_grow(n->_idx, i); } 284 285 // visited set accessors 286 void visited_clear() { _visited.Clear(); } 287 void visited_set(Node* n) { return _visited.set(bb_idx(n)); } 288 int visited_test(Node* n) { return _visited.test(bb_idx(n)); } 289 int visited_test_set(Node* n) { return _visited.test_set(bb_idx(n)); } 290 void post_visited_clear() { _post_visited.Clear(); } 291 void post_visited_set(Node* n) { return _post_visited.set(bb_idx(n)); } 292 int post_visited_test(Node* n) { return _post_visited.test(bb_idx(n)); } 293 294 // Ensure node_info contains element "i" 295 void grow_node_info(int i) { if (i >= _node_info.length()) _node_info.at_put_grow(i, SWNodeInfo::initial); } 296 297 // memory alignment for a node 298 int alignment(Node* n) { return _node_info.adr_at(bb_idx(n))->_alignment; } 299 void set_alignment(Node* n, int a) { int i = bb_idx(n); grow_node_info(i); _node_info.adr_at(i)->_alignment = a; } 300 301 // Max expression (DAG) depth from beginning of the block for each node 302 int depth(Node* n) { return _node_info.adr_at(bb_idx(n))->_depth; } 303 void set_depth(Node* n, int d) { int i = bb_idx(n); grow_node_info(i); _node_info.adr_at(i)->_depth = d; } 304 305 // vector element type 306 const Type* velt_type(Node* n) { return _node_info.adr_at(bb_idx(n))->_velt_type; } 307 BasicType velt_basic_type(Node* n) { return velt_type(n)->array_element_basic_type(); } 308 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; } 309 bool same_velt_type(Node* n1, Node* n2); 310 311 // my_pack 312 Node_List* my_pack(Node* n) { return !in_bb(n) ? NULL : _node_info.adr_at(bb_idx(n))->_my_pack; } 313 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; } 314 315 // methods 316 317 // Extract the superword level parallelism 318 void SLP_extract(); 319 // Find the adjacent memory references and create pack pairs for them. 320 void find_adjacent_refs(); 321 // Find a memory reference to align the loop induction variable to. 322 MemNode* find_align_to_ref(Node_List &memops); 323 // Calculate loop's iv adjustment for this memory ops. 324 int get_iv_adjustment(MemNode* mem); 325 // Can the preloop align the reference to position zero in the vector? 326 bool ref_is_alignable(SWPointer& p); 327 // Construct dependency graph. 328 void dependence_graph(); 329 // Return a memory slice (node list) in predecessor order starting at "start" 330 void mem_slice_preds(Node* start, Node* stop, GrowableArray<Node*> &preds); 331 // Can s1 and s2 be in a pack with s1 immediately preceding s2 and s1 aligned at "align" 332 bool stmts_can_pack(Node* s1, Node* s2, int align); 333 // Does s exist in a pack at position pos? 334 bool exists_at(Node* s, uint pos); 335 // Is s1 immediately before s2 in memory? 336 bool are_adjacent_refs(Node* s1, Node* s2); 337 // Are s1 and s2 similar? 338 bool isomorphic(Node* s1, Node* s2); 339 // Is there no data path from s1 to s2 or s2 to s1? 340 bool independent(Node* s1, Node* s2); 341 // Helper for independent 342 bool independent_path(Node* shallow, Node* deep, uint dp=0); 343 void set_alignment(Node* s1, Node* s2, int align); 344 int data_size(Node* s); 345 // Extend packset by following use->def and def->use links from pack members. 346 void extend_packlist(); 347 // Extend the packset by visiting operand definitions of nodes in pack p 348 bool follow_use_defs(Node_List* p); 349 // Extend the packset by visiting uses of nodes in pack p 350 bool follow_def_uses(Node_List* p); 351 // Estimate the savings from executing s1 and s2 as a pack 352 int est_savings(Node* s1, Node* s2); 353 int adjacent_profit(Node* s1, Node* s2); 354 int pack_cost(int ct); 355 int unpack_cost(int ct); 356 // Combine packs A and B with A.last == B.first into A.first..,A.last,B.second,..B.last 357 void combine_packs(); 358 // Construct the map from nodes to packs. 359 void construct_my_pack_map(); 360 // Remove packs that are not implemented or not profitable. 361 void filter_packs(); 362 // Adjust the memory graph for the packed operations 363 void schedule(); 364 // Remove "current" from its current position in the memory graph and insert 365 // it after the appropriate insert points (lip or uip); 366 void remove_and_insert(MemNode *current, MemNode *prev, MemNode *lip, Node *uip, Unique_Node_List &schd_before); 367 // Within a store pack, schedule stores together by moving out the sandwiched memory ops according 368 // to dependence info; and within a load pack, move loads down to the last executed load. 369 void co_locate_pack(Node_List* p); 370 // Convert packs into vector node operations 371 void output(); 372 // Create a vector operand for the nodes in pack p for operand: in(opd_idx) 373 Node* vector_opd(Node_List* p, int opd_idx); 374 // Can code be generated for pack p? 375 bool implemented(Node_List* p); 376 // For pack p, are all operands and all uses (with in the block) vector? 377 bool profitable(Node_List* p); 378 // If a use of pack p is not a vector use, then replace the use with an extract operation. 379 void insert_extracts(Node_List* p); 380 // Is use->in(u_idx) a vector use? 381 bool is_vector_use(Node* use, int u_idx); 382 // Construct reverse postorder list of block members 383 bool construct_bb(); 384 // Initialize per node info 385 void initialize_bb(); 386 // Insert n into block after pos 387 void bb_insert_after(Node* n, int pos); 388 // Compute max depth for expressions from beginning of block 389 void compute_max_depth(); 390 // Compute necessary vector element type for expressions 391 void compute_vector_element_type(); 392 // Are s1 and s2 in a pack pair and ordered as s1,s2? 393 bool in_packset(Node* s1, Node* s2); 394 // Is s in pack p? 395 Node_List* in_pack(Node* s, Node_List* p); 396 // Remove the pack at position pos in the packset 397 void remove_pack_at(int pos); 398 // Return the node executed first in pack p. 399 Node* executed_first(Node_List* p); 400 // Return the node executed last in pack p. 401 Node* executed_last(Node_List* p); 402 // Alignment within a vector memory reference 403 int memory_alignment(MemNode* s, int iv_adjust); 404 // (Start, end] half-open range defining which operands are vector 405 void vector_opd_range(Node* n, uint* start, uint* end); 406 // Smallest type containing range of values 407 const Type* container_type(Node* n); 408 // Adjust pre-loop limit so that in main loop, a load/store reference 409 // to align_to_ref will be a position zero in the vector. 410 void align_initial_loop_index(MemNode* align_to_ref); 411 // Find pre loop end from main loop. Returns null if none. 412 CountedLoopEndNode* get_pre_loop_end(CountedLoopNode *cl); 413 // Is the use of d1 in u1 at the same operand position as d2 in u2? 414 bool opnd_positions_match(Node* d1, Node* u1, Node* d2, Node* u2); 415 void init(); 416 417 // print methods 418 void print_packset(); 419 void print_pack(Node_List* p); 420 void print_bb(); 421 void print_stmt(Node* s); 422 char* blank(uint depth); 423 }; 424 425 426 //------------------------------SWPointer--------------------------- 427 // Information about an address for dependence checking and vector alignment 428 class SWPointer VALUE_OBJ_CLASS_SPEC { 429 protected: 430 MemNode* _mem; // My memory reference node 431 SuperWord* _slp; // SuperWord class 432 433 Node* _base; // NULL if unsafe nonheap reference 434 Node* _adr; // address pointer 435 jint _scale; // multiplier for iv (in bytes), 0 if no loop iv 436 jint _offset; // constant offset (in bytes) 437 Node* _invar; // invariant offset (in bytes), NULL if none 438 bool _negate_invar; // if true then use: (0 - _invar) 439 440 PhaseIdealLoop* phase() { return _slp->phase(); } 441 IdealLoopTree* lpt() { return _slp->lpt(); } 442 PhiNode* iv() { return _slp->iv(); } // Induction var 443 444 bool invariant(Node* n) { 445 Node *n_c = phase()->get_ctrl(n); 446 return !lpt()->is_member(phase()->get_loop(n_c)); 447 } 448 449 // Match: k*iv + offset 450 bool scaled_iv_plus_offset(Node* n); 451 // Match: k*iv where k is a constant that's not zero 452 bool scaled_iv(Node* n); 453 // Match: offset is (k [+/- invariant]) 454 bool offset_plus_k(Node* n, bool negate = false); 455 456 public: 457 enum CMP { 458 Less = 1, 459 Greater = 2, 460 Equal = 4, 461 NotEqual = (Less | Greater), 462 NotComparable = (Less | Greater | Equal) 463 }; 464 465 SWPointer(MemNode* mem, SuperWord* slp); 466 // Following is used to create a temporary object during 467 // the pattern match of an address expression. 468 SWPointer(SWPointer* p); 469 470 bool valid() { return _adr != NULL; } 471 bool has_iv() { return _scale != 0; } 472 473 Node* base() { return _base; } 474 Node* adr() { return _adr; } 475 MemNode* mem() { return _mem; } 476 int scale_in_bytes() { return _scale; } 477 Node* invar() { return _invar; } 478 bool negate_invar() { return _negate_invar; } 479 int offset_in_bytes() { return _offset; } 480 int memory_size() { return _mem->memory_size(); } 481 482 // Comparable? 483 int cmp(SWPointer& q) { 484 if (valid() && q.valid() && 485 (_adr == q._adr || _base == _adr && q._base == q._adr) && 486 _scale == q._scale && 487 _invar == q._invar && 488 _negate_invar == q._negate_invar) { 489 bool overlap = q._offset < _offset + memory_size() && 490 _offset < q._offset + q.memory_size(); 491 return overlap ? Equal : (_offset < q._offset ? Less : Greater); 492 } else { 493 return NotComparable; 494 } 495 } 496 497 bool not_equal(SWPointer& q) { return not_equal(cmp(q)); } 498 bool equal(SWPointer& q) { return equal(cmp(q)); } 499 bool comparable(SWPointer& q) { return comparable(cmp(q)); } 500 static bool not_equal(int cmp) { return cmp <= NotEqual; } 501 static bool equal(int cmp) { return cmp == Equal; } 502 static bool comparable(int cmp) { return cmp < NotComparable; } 503 504 void print(); 505 }; 506 507 508 //------------------------------OrderedPair--------------------------- 509 // Ordered pair of Node*. 510 class OrderedPair VALUE_OBJ_CLASS_SPEC { 511 protected: 512 Node* _p1; 513 Node* _p2; 514 public: 515 OrderedPair() : _p1(NULL), _p2(NULL) {} 516 OrderedPair(Node* p1, Node* p2) { 517 if (p1->_idx < p2->_idx) { 518 _p1 = p1; _p2 = p2; 519 } else { 520 _p1 = p2; _p2 = p1; 521 } 522 } 523 524 bool operator==(const OrderedPair &rhs) { 525 return _p1 == rhs._p1 && _p2 == rhs._p2; 526 } 527 void print() { tty->print(" (%d, %d)", _p1->_idx, _p2->_idx); } 528 529 static const OrderedPair initial; 530 }; 531 532 #endif // SHARE_VM_OPTO_SUPERWORD_HPP --- EOF ---