1 /*
2 * Copyright (c) 2007, 2018, 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 #include "libadt/dict.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 {
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 {
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 class SuperWord;
205 class CMoveKit {
206 friend class SuperWord;
207 private:
208 SuperWord* _sw;
209 Dict* _dict;
210 CMoveKit(Arena* a, SuperWord* sw) : _sw(sw) {_dict = new Dict(cmpkey, hashkey, a);}
211 void* _2p(Node* key) const { return (void*)(intptr_t)key; } // 2 conversion functions to make gcc happy
212 Dict* dict() const { return _dict; }
213 void map(Node* key, Node_List* val) { assert(_dict->operator[](_2p(key)) == NULL, "key existed"); _dict->Insert(_2p(key), (void*)val); }
214 void unmap(Node* key) { _dict->Delete(_2p(key)); }
215 Node_List* pack(Node* key) const { return (Node_List*)_dict->operator[](_2p(key)); }
216 Node* is_Bool_candidate(Node* nd) const; // if it is the right candidate return corresponding CMove* ,
217 Node* is_CmpD_candidate(Node* nd) const; // otherwise return NULL
218 Node_List* make_cmovevd_pack(Node_List* cmovd_pk);
219 bool test_cmpd_pack(Node_List* cmpd_pk, Node_List* cmovd_pk);
220 };//class CMoveKit
221
222 // JVMCI: OrderedPair is moved up to deal with compilation issues on Windows
223 //------------------------------OrderedPair---------------------------
224 // Ordered pair of Node*.
225 class OrderedPair {
226 protected:
227 Node* _p1;
228 Node* _p2;
229 public:
230 OrderedPair() : _p1(NULL), _p2(NULL) {}
231 OrderedPair(Node* p1, Node* p2) {
232 if (p1->_idx < p2->_idx) {
233 _p1 = p1; _p2 = p2;
234 } else {
235 _p1 = p2; _p2 = p1;
236 }
237 }
238
239 bool operator==(const OrderedPair &rhs) {
240 return _p1 == rhs._p1 && _p2 == rhs._p2;
241 }
242 void print() { tty->print(" (%d, %d)", _p1->_idx, _p2->_idx); }
243
244 static const OrderedPair initial;
245 };
246
247 // -----------------------------SuperWord---------------------------------
248 // Transforms scalar operations into packed (superword) operations.
249 class SuperWord : public ResourceObj {
250 friend class SWPointer;
251 friend class CMoveKit;
252 private:
253 PhaseIdealLoop* _phase;
254 Arena* _arena;
255 PhaseIterGVN &_igvn;
256
257 enum consts { top_align = -1, bottom_align = -666 };
258
259 GrowableArray<Node_List*> _packset; // Packs for the current block
260
261 GrowableArray<int> _bb_idx; // Map from Node _idx to index within block
262
263 GrowableArray<Node*> _block; // Nodes in current block
264 GrowableArray<Node*> _post_block; // Nodes in post loop block
265 GrowableArray<Node*> _data_entry; // Nodes with all inputs from outside
266 GrowableArray<Node*> _mem_slice_head; // Memory slice head nodes
267 GrowableArray<Node*> _mem_slice_tail; // Memory slice tail nodes
268 GrowableArray<Node*> _iteration_first; // nodes in the generation that has deps from phi
269 GrowableArray<Node*> _iteration_last; // nodes in the generation that has deps to phi
270 GrowableArray<SWNodeInfo> _node_info; // Info needed per node
271 CloneMap& _clone_map; // map of nodes created in cloning
272 CMoveKit _cmovev_kit; // support for vectorization of CMov
273 MemNode* _align_to_ref; // Memory reference that pre-loop will align to
274
275 GrowableArray<OrderedPair> _disjoint_ptrs; // runtime disambiguated pointer pairs
276
277 DepGraph _dg; // Dependence graph
278
279 // Scratch pads
280 VectorSet _visited; // Visited set
281 VectorSet _post_visited; // Post-visited set
282 Node_Stack _n_idx_list; // List of (node,index) pairs
283 GrowableArray<Node*> _nlist; // List of nodes
284 GrowableArray<Node*> _stk; // Stack of nodes
285
286 public:
287 SuperWord(PhaseIdealLoop* phase);
288
289 void transform_loop(IdealLoopTree* lpt, bool do_optimization);
290
291 void unrolling_analysis(int &local_loop_unroll_factor);
292
293 // Accessors for SWPointer
294 PhaseIdealLoop* phase() { return _phase; }
295 IdealLoopTree* lpt() { return _lpt; }
296 PhiNode* iv() { return _iv; }
297
298 bool early_return() { return _early_return; }
299
300 #ifndef PRODUCT
301 bool is_debug() { return _vector_loop_debug > 0; }
302 bool is_trace_alignment() { return (_vector_loop_debug & 2) > 0; }
303 bool is_trace_mem_slice() { return (_vector_loop_debug & 4) > 0; }
304 bool is_trace_loop() { return (_vector_loop_debug & 8) > 0; }
305 bool is_trace_adjacent() { return (_vector_loop_debug & 16) > 0; }
306 bool is_trace_cmov() { return (_vector_loop_debug & 32) > 0; }
307 bool is_trace_loop_reverse() { return (_vector_loop_debug & 64) > 0; }
308 #endif
309 bool do_vector_loop() { return _do_vector_loop; }
310 bool do_reserve_copy() { return _do_reserve_copy; }
311 private:
312 IdealLoopTree* _lpt; // Current loop tree node
313 LoopNode* _lp; // Current LoopNode
314 Node* _bb; // Current basic block
315 PhiNode* _iv; // Induction var
316 bool _race_possible; // In cases where SDMU is true
317 bool _early_return; // True if we do not initialize
318 bool _do_vector_loop; // whether to do vectorization/simd style
319 bool _do_reserve_copy; // do reserve copy of the graph(loop) before final modification in output
320 int _num_work_vecs; // Number of non memory vector operations
321 int _num_reductions; // Number of reduction expressions applied
322 int _ii_first; // generation with direct deps from mem phi
323 int _ii_last; // generation with direct deps to mem phi
324 GrowableArray<int> _ii_order;
325 #ifndef PRODUCT
326 uintx _vector_loop_debug; // provide more printing in debug mode
327 #endif
328
329 // Accessors
330 Arena* arena() { return _arena; }
331
332 Node* bb() { return _bb; }
333 void set_bb(Node* bb) { _bb = bb; }
334
335 void set_lpt(IdealLoopTree* lpt) { _lpt = lpt; }
336
337 LoopNode* lp() { return _lp; }
338 void set_lp(LoopNode* lp) { _lp = lp;
339 _iv = lp->as_CountedLoop()->phi()->as_Phi(); }
340 int iv_stride() { return lp()->as_CountedLoop()->stride_con(); }
341
342 int vector_width(Node* n) {
343 BasicType bt = velt_basic_type(n);
344 return MIN2(ABS(iv_stride()), Matcher::max_vector_size(bt));
345 }
346 int vector_width_in_bytes(Node* n) {
347 BasicType bt = velt_basic_type(n);
348 return vector_width(n)*type2aelembytes(bt);
349 }
350 MemNode* align_to_ref() { return _align_to_ref; }
351 void set_align_to_ref(MemNode* m) { _align_to_ref = m; }
352
353 Node* ctrl(Node* n) const { return _phase->has_ctrl(n) ? _phase->get_ctrl(n) : n; }
354
355 // block accessors
356 bool in_bb(Node* n) { return n != NULL && n->outcnt() > 0 && ctrl(n) == _bb; }
357 int bb_idx(Node* n) { assert(in_bb(n), "must be"); return _bb_idx.at(n->_idx); }
358 void set_bb_idx(Node* n, int i) { _bb_idx.at_put_grow(n->_idx, i); }
359
360 // visited set accessors
361 void visited_clear() { _visited.Clear(); }
362 void visited_set(Node* n) { return _visited.set(bb_idx(n)); }
363 int visited_test(Node* n) { return _visited.test(bb_idx(n)); }
364 int visited_test_set(Node* n) { return _visited.test_set(bb_idx(n)); }
365 void post_visited_clear() { _post_visited.Clear(); }
366 void post_visited_set(Node* n) { return _post_visited.set(bb_idx(n)); }
367 int post_visited_test(Node* n) { return _post_visited.test(bb_idx(n)); }
368
369 // Ensure node_info contains element "i"
370 void grow_node_info(int i) { if (i >= _node_info.length()) _node_info.at_put_grow(i, SWNodeInfo::initial); }
371
372 // memory alignment for a node
373 int alignment(Node* n) { return _node_info.adr_at(bb_idx(n))->_alignment; }
374 void set_alignment(Node* n, int a) { int i = bb_idx(n); grow_node_info(i); _node_info.adr_at(i)->_alignment = a; }
375
376 // Max expression (DAG) depth from beginning of the block for each node
377 int depth(Node* n) { return _node_info.adr_at(bb_idx(n))->_depth; }
378 void set_depth(Node* n, int d) { int i = bb_idx(n); grow_node_info(i); _node_info.adr_at(i)->_depth = d; }
379
380 // vector element type
381 const Type* velt_type(Node* n) { return _node_info.adr_at(bb_idx(n))->_velt_type; }
382 BasicType velt_basic_type(Node* n) { return velt_type(n)->array_element_basic_type(); }
383 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; }
384 bool same_velt_type(Node* n1, Node* n2);
385
386 // my_pack
387 Node_List* my_pack(Node* n) { return !in_bb(n) ? NULL : _node_info.adr_at(bb_idx(n))->_my_pack; }
388 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; }
389 // is pack good for converting into one vector node replacing 12 nodes of Cmp, Bool, CMov
390 bool is_cmov_pack(Node_List* p);
391 bool is_cmov_pack_internal_node(Node_List* p, Node* nd) { return is_cmov_pack(p) && !nd->is_CMove(); }
392 // For pack p, are all idx operands the same?
393 bool same_inputs(Node_List* p, int idx);
394 // CloneMap utilities
395 bool same_origin_idx(Node* a, Node* b) const;
396 bool same_generation(Node* a, Node* b) const;
397
398 // methods
399
400 // Extract the superword level parallelism
401 void SLP_extract();
402 // Find the adjacent memory references and create pack pairs for them.
403 void find_adjacent_refs();
404 // Tracing support
405 #ifndef PRODUCT
406 void find_adjacent_refs_trace_1(Node* best_align_to_mem_ref, int best_iv_adjustment);
407 void print_loop(bool whole);
408 #endif
409 // Find a memory reference to align the loop induction variable to.
410 MemNode* find_align_to_ref(Node_List &memops);
411 // Calculate loop's iv adjustment for this memory ops.
412 int get_iv_adjustment(MemNode* mem);
413 // Can the preloop align the reference to position zero in the vector?
414 bool ref_is_alignable(SWPointer& p);
415 // rebuild the graph so all loads in different iterations of cloned loop become dependant on phi node (in _do_vector_loop only)
416 bool hoist_loads_in_graph();
417 // Test whether MemNode::Memory dependency to the same load but in the first iteration of this loop is coming from memory phi
418 // Return false if failed
419 Node* find_phi_for_mem_dep(LoadNode* ld);
420 // Return same node but from the first generation. Return 0, if not found
421 Node* first_node(Node* nd);
422 // Return same node as this but from the last generation. Return 0, if not found
423 Node* last_node(Node* n);
424 // Mark nodes belonging to first and last generation
425 // returns first generation index or -1 if vectorization/simd is impossible
426 int mark_generations();
427 // swapping inputs of commutative instruction (Add or Mul)
428 bool fix_commutative_inputs(Node* gold, Node* fix);
429 // make packs forcefully (in _do_vector_loop only)
430 bool pack_parallel();
431 // Construct dependency graph.
432 void dependence_graph();
433 // Return a memory slice (node list) in predecessor order starting at "start"
434 void mem_slice_preds(Node* start, Node* stop, GrowableArray<Node*> &preds);
435 // Can s1 and s2 be in a pack with s1 immediately preceding s2 and s1 aligned at "align"
436 bool stmts_can_pack(Node* s1, Node* s2, int align);
437 // Does s exist in a pack at position pos?
438 bool exists_at(Node* s, uint pos);
439 // Is s1 immediately before s2 in memory?
440 bool are_adjacent_refs(Node* s1, Node* s2);
441 // Are s1 and s2 similar?
442 bool isomorphic(Node* s1, Node* s2);
443 // Is there no data path from s1 to s2 or s2 to s1?
444 bool independent(Node* s1, Node* s2);
445 // For a node pair (s1, s2) which is isomorphic and independent,
446 // do s1 and s2 have similar input edges?
447 bool have_similar_inputs(Node* s1, Node* s2);
448 // Is there a data path between s1 and s2 and both are reductions?
449 bool reduction(Node* s1, Node* s2);
450 // Helper for independent
451 bool independent_path(Node* shallow, Node* deep, uint dp=0);
452 void set_alignment(Node* s1, Node* s2, int align);
453 int data_size(Node* s);
454 // Extend packset by following use->def and def->use links from pack members.
455 void extend_packlist();
456 // Extend the packset by visiting operand definitions of nodes in pack p
457 bool follow_use_defs(Node_List* p);
458 // Extend the packset by visiting uses of nodes in pack p
459 bool follow_def_uses(Node_List* p);
460 // For extended packsets, ordinally arrange uses packset by major component
461 void order_def_uses(Node_List* p);
462 // Estimate the savings from executing s1 and s2 as a pack
463 int est_savings(Node* s1, Node* s2);
464 int adjacent_profit(Node* s1, Node* s2);
465 int pack_cost(int ct);
466 int unpack_cost(int ct);
467 // Combine packs A and B with A.last == B.first into A.first..,A.last,B.second,..B.last
468 void combine_packs();
469 // Construct the map from nodes to packs.
470 void construct_my_pack_map();
471 // Remove packs that are not implemented or not profitable.
472 void filter_packs();
473 // Merge CMoveD into new vector-nodes
474 void merge_packs_to_cmovd();
475 // Adjust the memory graph for the packed operations
476 void schedule();
477 // Remove "current" from its current position in the memory graph and insert
478 // it after the appropriate insert points (lip or uip);
479 void remove_and_insert(MemNode *current, MemNode *prev, MemNode *lip, Node *uip, Unique_Node_List &schd_before);
480 // Within a store pack, schedule stores together by moving out the sandwiched memory ops according
481 // to dependence info; and within a load pack, move loads down to the last executed load.
482 void co_locate_pack(Node_List* p);
483 // Convert packs into vector node operations
484 void output();
485 // Create a vector operand for the nodes in pack p for operand: in(opd_idx)
486 Node* vector_opd(Node_List* p, int opd_idx);
487 // Can code be generated for pack p?
488 bool implemented(Node_List* p);
489 // For pack p, are all operands and all uses (with in the block) vector?
490 bool profitable(Node_List* p);
491 // If a use of pack p is not a vector use, then replace the use with an extract operation.
492 void insert_extracts(Node_List* p);
493 // Is use->in(u_idx) a vector use?
494 bool is_vector_use(Node* use, int u_idx);
495 // Construct reverse postorder list of block members
496 bool construct_bb();
497 // Initialize per node info
498 void initialize_bb();
499 // Insert n into block after pos
500 void bb_insert_after(Node* n, int pos);
501 // Compute max depth for expressions from beginning of block
502 void compute_max_depth();
503 // Compute necessary vector element type for expressions
504 void compute_vector_element_type();
505 // Are s1 and s2 in a pack pair and ordered as s1,s2?
506 bool in_packset(Node* s1, Node* s2);
507 // Is s in pack p?
508 Node_List* in_pack(Node* s, Node_List* p);
509 // Remove the pack at position pos in the packset
510 void remove_pack_at(int pos);
511 // Return the node executed first in pack p.
512 Node* executed_first(Node_List* p);
513 // Return the node executed last in pack p.
514 Node* executed_last(Node_List* p);
515 static LoadNode::ControlDependency control_dependency(Node_List* p);
516 // Alignment within a vector memory reference
517 int memory_alignment(MemNode* s, int iv_adjust);
518 // (Start, end] half-open range defining which operands are vector
519 void vector_opd_range(Node* n, uint* start, uint* end);
520 // Smallest type containing range of values
521 const Type* container_type(Node* n);
522 // Adjust pre-loop limit so that in main loop, a load/store reference
523 // to align_to_ref will be a position zero in the vector.
524 void align_initial_loop_index(MemNode* align_to_ref);
525 // Find pre loop end from main loop. Returns null if none.
526 CountedLoopEndNode* get_pre_loop_end(CountedLoopNode *cl);
527 // Is the use of d1 in u1 at the same operand position as d2 in u2?
528 bool opnd_positions_match(Node* d1, Node* u1, Node* d2, Node* u2);
529 void init();
530 // clean up some basic structures - used if the ideal graph was rebuilt
531 void restart();
532
533 // print methods
534 void print_packset();
535 void print_pack(Node_List* p);
536 void print_bb();
537 void print_stmt(Node* s);
538 char* blank(uint depth);
539
540 void packset_sort(int n);
541 };
542
543
544
545 //------------------------------SWPointer---------------------------
546 // Information about an address for dependence checking and vector alignment
547 class SWPointer {
548 protected:
549 MemNode* _mem; // My memory reference node
550 SuperWord* _slp; // SuperWord class
551
552 Node* _base; // NULL if unsafe nonheap reference
553 Node* _adr; // address pointer
554 jint _scale; // multiplier for iv (in bytes), 0 if no loop iv
555 jint _offset; // constant offset (in bytes)
556 Node* _invar; // invariant offset (in bytes), NULL if none
557 bool _negate_invar; // if true then use: (0 - _invar)
558 Node_Stack* _nstack; // stack used to record a swpointer trace of variants
559 bool _analyze_only; // Used in loop unrolling only for swpointer trace
560 uint _stack_idx; // Used in loop unrolling only for swpointer trace
561
562 PhaseIdealLoop* phase() { return _slp->phase(); }
563 IdealLoopTree* lpt() { return _slp->lpt(); }
564 PhiNode* iv() { return _slp->iv(); } // Induction var
565
566 bool invariant(Node* n);
567
568 // Match: k*iv + offset
569 bool scaled_iv_plus_offset(Node* n);
570 // Match: k*iv where k is a constant that's not zero
571 bool scaled_iv(Node* n);
572 // Match: offset is (k [+/- invariant])
573 bool offset_plus_k(Node* n, bool negate = false);
574
575 public:
576 enum CMP {
577 Less = 1,
578 Greater = 2,
579 Equal = 4,
580 NotEqual = (Less | Greater),
581 NotComparable = (Less | Greater | Equal)
582 };
583
584 SWPointer(MemNode* mem, SuperWord* slp, Node_Stack *nstack, bool analyze_only);
585 // Following is used to create a temporary object during
586 // the pattern match of an address expression.
587 SWPointer(SWPointer* p);
588
589 bool valid() { return _adr != NULL; }
590 bool has_iv() { return _scale != 0; }
591
592 Node* base() { return _base; }
593 Node* adr() { return _adr; }
594 MemNode* mem() { return _mem; }
595 int scale_in_bytes() { return _scale; }
596 Node* invar() { return _invar; }
597 bool negate_invar() { return _negate_invar; }
598 int offset_in_bytes() { return _offset; }
599 int memory_size() { return _mem->memory_size(); }
600 Node_Stack* node_stack() { return _nstack; }
601
602 // Comparable?
603 int cmp(SWPointer& q) {
604 if (valid() && q.valid() &&
605 (_adr == q._adr || (_base == _adr && q._base == q._adr)) &&
606 _scale == q._scale &&
607 _invar == q._invar &&
608 _negate_invar == q._negate_invar) {
609 bool overlap = q._offset < _offset + memory_size() &&
610 _offset < q._offset + q.memory_size();
611 return overlap ? Equal : (_offset < q._offset ? Less : Greater);
612 } else {
613 return NotComparable;
614 }
615 }
616
617 bool not_equal(SWPointer& q) { return not_equal(cmp(q)); }
618 bool equal(SWPointer& q) { return equal(cmp(q)); }
619 bool comparable(SWPointer& q) { return comparable(cmp(q)); }
620 static bool not_equal(int cmp) { return cmp <= NotEqual; }
621 static bool equal(int cmp) { return cmp == Equal; }
622 static bool comparable(int cmp) { return cmp < NotComparable; }
623
624 void print();
625
626 #ifndef PRODUCT
627 class Tracer {
628 friend class SuperWord;
629 friend class SWPointer;
630 SuperWord* _slp;
631 static int _depth;
632 int _depth_save;
633 void print_depth();
634 int depth() const { return _depth; }
635 void set_depth(int d) { _depth = d; }
636 void inc_depth() { _depth++;}
637 void dec_depth() { if (_depth > 0) _depth--;}
638 void store_depth() {_depth_save = _depth;}
639 void restore_depth() {_depth = _depth_save;}
640
641 class Depth {
642 friend class Tracer;
643 friend class SWPointer;
644 friend class SuperWord;
645 Depth() { ++_depth; }
646 Depth(int x) { _depth = 0; }
647 ~Depth() { if (_depth > 0) --_depth;}
648 };
649 Tracer (SuperWord* slp) : _slp(slp) {}
650
651 // tracing functions
652 void ctor_1(Node* mem);
653 void ctor_2(Node* adr);
654 void ctor_3(Node* adr, int i);
655 void ctor_4(Node* adr, int i);
656 void ctor_5(Node* adr, Node* base, int i);
657 void ctor_6(Node* mem);
658
659 void invariant_1(Node *n, Node *n_c);
660
661 void scaled_iv_plus_offset_1(Node* n);
662 void scaled_iv_plus_offset_2(Node* n);
663 void scaled_iv_plus_offset_3(Node* n);
664 void scaled_iv_plus_offset_4(Node* n);
665 void scaled_iv_plus_offset_5(Node* n);
666 void scaled_iv_plus_offset_6(Node* n);
667 void scaled_iv_plus_offset_7(Node* n);
668 void scaled_iv_plus_offset_8(Node* n);
669
670 void scaled_iv_1(Node* n);
671 void scaled_iv_2(Node* n, int scale);
672 void scaled_iv_3(Node* n, int scale);
673 void scaled_iv_4(Node* n, int scale);
674 void scaled_iv_5(Node* n, int scale);
675 void scaled_iv_6(Node* n, int scale);
676 void scaled_iv_7(Node* n);
677 void scaled_iv_8(Node* n, SWPointer* tmp);
678 void scaled_iv_9(Node* n, int _scale, int _offset, int mult);
679 void scaled_iv_10(Node* n);
680
681 void offset_plus_k_1(Node* n);
682 void offset_plus_k_2(Node* n, int _offset);
683 void offset_plus_k_3(Node* n, int _offset);
684 void offset_plus_k_4(Node* n);
685 void offset_plus_k_5(Node* n, Node* _invar);
686 void offset_plus_k_6(Node* n, Node* _invar, bool _negate_invar, int _offset);
687 void offset_plus_k_7(Node* n, Node* _invar, bool _negate_invar, int _offset);
688 void offset_plus_k_8(Node* n, Node* _invar, bool _negate_invar, int _offset);
689 void offset_plus_k_9(Node* n, Node* _invar, bool _negate_invar, int _offset);
690 void offset_plus_k_10(Node* n, Node* _invar, bool _negate_invar, int _offset);
691 void offset_plus_k_11(Node* n);
692
693 } _tracer;//TRacer;
694 #endif
695 };
696
697 #endif // SHARE_VM_OPTO_SUPERWORD_HPP