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/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 Amarasighe
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
222 GrowableArray<SWNodeInfo> _node_info; // Info needed per node
223
224 MemNode* _align_to_ref; // Memory reference that pre-loop will align to
225
226 GrowableArray<OrderedPair> _disjoint_ptrs; // runtime disambiguated pointer pairs
227
228 DepGraph _dg; // Dependence graph
229
230 // Scratch pads
231 VectorSet _visited; // Visited set
232 VectorSet _post_visited; // Post-visited set
233 Node_Stack _n_idx_list; // List of (node,index) pairs
234 GrowableArray<Node*> _nlist; // List of nodes
235 GrowableArray<Node*> _stk; // Stack of nodes
236
237 public:
238 SuperWord(PhaseIdealLoop* phase);
239
240 void transform_loop(IdealLoopTree* lpt, bool do_optimization);
241
242 // Accessors for SWPointer
243 PhaseIdealLoop* phase() { return _phase; }
244 IdealLoopTree* lpt() { return _lpt; }
245 PhiNode* iv() { return _iv; }
246 bool early_return() { return _early_return; }
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 bool _race_possible; // In cases where SDMU is true
254 bool _early_return; // True if we do not initialize
255
256 // Accessors
257 Arena* arena() { return _arena; }
258
259 Node* bb() { return _bb; }
260 void set_bb(Node* bb) { _bb = bb; }
261
262 void set_lpt(IdealLoopTree* lpt) { _lpt = lpt; }
263
264 LoopNode* lp() { return _lp; }
265 void set_lp(LoopNode* lp) { _lp = lp;
266 _iv = lp->as_CountedLoop()->phi()->as_Phi(); }
267 int iv_stride() { return lp()->as_CountedLoop()->stride_con(); }
268
269 int vector_width(Node* n) {
270 BasicType bt = velt_basic_type(n);
271 return MIN2(ABS(iv_stride()), Matcher::max_vector_size(bt));
272 }
273 int vector_width_in_bytes(Node* n) {
274 BasicType bt = velt_basic_type(n);
275 return vector_width(n)*type2aelembytes(bt);
276 }
277 MemNode* align_to_ref() { return _align_to_ref; }
278 void set_align_to_ref(MemNode* m) { _align_to_ref = m; }
279
280 Node* ctrl(Node* n) const { return _phase->has_ctrl(n) ? _phase->get_ctrl(n) : n; }
281
282 // block accessors
283 bool in_bb(Node* n) { return n != NULL && n->outcnt() > 0 && ctrl(n) == _bb; }
284 int bb_idx(Node* n) { assert(in_bb(n), "must be"); return _bb_idx.at(n->_idx); }
285 void set_bb_idx(Node* n, int i) { _bb_idx.at_put_grow(n->_idx, i); }
286
287 // visited set accessors
288 void visited_clear() { _visited.Clear(); }
289 void visited_set(Node* n) { return _visited.set(bb_idx(n)); }
290 int visited_test(Node* n) { return _visited.test(bb_idx(n)); }
291 int visited_test_set(Node* n) { return _visited.test_set(bb_idx(n)); }
292 void post_visited_clear() { _post_visited.Clear(); }
293 void post_visited_set(Node* n) { return _post_visited.set(bb_idx(n)); }
294 int post_visited_test(Node* n) { return _post_visited.test(bb_idx(n)); }
295
296 // Ensure node_info contains element "i"
297 void grow_node_info(int i) { if (i >= _node_info.length()) _node_info.at_put_grow(i, SWNodeInfo::initial); }
298
299 // memory alignment for a node
300 int alignment(Node* n) { return _node_info.adr_at(bb_idx(n))->_alignment; }
301 void set_alignment(Node* n, int a) { int i = bb_idx(n); grow_node_info(i); _node_info.adr_at(i)->_alignment = a; }
302
303 // Max expression (DAG) depth from beginning of the block for each node
304 int depth(Node* n) { return _node_info.adr_at(bb_idx(n))->_depth; }
305 void set_depth(Node* n, int d) { int i = bb_idx(n); grow_node_info(i); _node_info.adr_at(i)->_depth = d; }
306
307 // vector element type
308 const Type* velt_type(Node* n) { return _node_info.adr_at(bb_idx(n))->_velt_type; }
309 BasicType velt_basic_type(Node* n) { return velt_type(n)->array_element_basic_type(); }
310 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; }
311 bool same_velt_type(Node* n1, Node* n2);
312
313 // my_pack
314 Node_List* my_pack(Node* n) { return !in_bb(n) ? NULL : _node_info.adr_at(bb_idx(n))->_my_pack; }
315 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; }
316
317 // methods
318
319 // Extract the superword level parallelism
320 void SLP_extract();
321 // Find the adjacent memory references and create pack pairs for them.
322 void find_adjacent_refs();
323 // Find a memory reference to align the loop induction variable to.
324 MemNode* find_align_to_ref(Node_List &memops);
325 // Calculate loop's iv adjustment for this memory ops.
326 int get_iv_adjustment(MemNode* mem);
327 // Can the preloop align the reference to position zero in the vector?
328 bool ref_is_alignable(SWPointer& p);
329 // Construct dependency graph.
330 void dependence_graph();
331 // Return a memory slice (node list) in predecessor order starting at "start"
332 void mem_slice_preds(Node* start, Node* stop, GrowableArray<Node*> &preds);
333 // Can s1 and s2 be in a pack with s1 immediately preceding s2 and s1 aligned at "align"
334 bool stmts_can_pack(Node* s1, Node* s2, int align);
335 // Does s exist in a pack at position pos?
336 bool exists_at(Node* s, uint pos);
337 // Is s1 immediately before s2 in memory?
338 bool are_adjacent_refs(Node* s1, Node* s2);
339 // Are s1 and s2 similar?
340 bool isomorphic(Node* s1, Node* s2);
341 // Is there no data path from s1 to s2 or s2 to s1?
342 bool independent(Node* s1, Node* s2);
343 // Is there a data path between s1 and s2 and both are reductions?
344 bool reduction(Node* s1, Node* s2);
345 // Helper for independent
346 bool independent_path(Node* shallow, Node* deep, uint dp=0);
347 void set_alignment(Node* s1, Node* s2, int align);
348 int data_size(Node* s);
349 // Extend packset by following use->def and def->use links from pack members.
350 void extend_packlist();
351 // Extend the packset by visiting operand definitions of nodes in pack p
352 bool follow_use_defs(Node_List* p);
353 // Extend the packset by visiting uses of nodes in pack p
354 bool follow_def_uses(Node_List* p);
355 // For extended packsets, ordinally arrange uses packset by major component
356 void order_def_uses(Node_List* p);
357 // Estimate the savings from executing s1 and s2 as a pack
358 int est_savings(Node* s1, Node* s2);
359 int adjacent_profit(Node* s1, Node* s2);
360 int pack_cost(int ct);
361 int unpack_cost(int ct);
362 // Combine packs A and B with A.last == B.first into A.first..,A.last,B.second,..B.last
363 void combine_packs();
364 // Construct the map from nodes to packs.
365 void construct_my_pack_map();
366 // Remove packs that are not implemented or not profitable.
367 void filter_packs();
368 // Adjust the memory graph for the packed operations
369 void schedule();
370 // Remove "current" from its current position in the memory graph and insert
371 // it after the appropriate insert points (lip or uip);
372 void remove_and_insert(MemNode *current, MemNode *prev, MemNode *lip, Node *uip, Unique_Node_List &schd_before);
373 // Within a store pack, schedule stores together by moving out the sandwiched memory ops according
374 // to dependence info; and within a load pack, move loads down to the last executed load.
375 void co_locate_pack(Node_List* p);
376 // Convert packs into vector node operations
377 void output();
378 // Create a vector operand for the nodes in pack p for operand: in(opd_idx)
379 Node* vector_opd(Node_List* p, int opd_idx);
380 // Can code be generated for pack p?
381 bool implemented(Node_List* p);
382 // For pack p, are all operands and all uses (with in the block) vector?
383 bool profitable(Node_List* p);
384 // If a use of pack p is not a vector use, then replace the use with an extract operation.
385 void insert_extracts(Node_List* p);
386 // Is use->in(u_idx) a vector use?
387 bool is_vector_use(Node* use, int u_idx);
388 // Construct reverse postorder list of block members
389 bool construct_bb();
390 // Initialize per node info
391 void initialize_bb();
392 // Insert n into block after pos
393 void bb_insert_after(Node* n, int pos);
394 // Compute max depth for expressions from beginning of block
395 void compute_max_depth();
396 // Compute necessary vector element type for expressions
397 void compute_vector_element_type();
398 // Are s1 and s2 in a pack pair and ordered as s1,s2?
399 bool in_packset(Node* s1, Node* s2);
400 // Is s in pack p?
401 Node_List* in_pack(Node* s, Node_List* p);
402 // Remove the pack at position pos in the packset
403 void remove_pack_at(int pos);
404 // Return the node executed first in pack p.
405 Node* executed_first(Node_List* p);
406 // Return the node executed last in pack p.
407 Node* executed_last(Node_List* p);
408 // Alignment within a vector memory reference
409 int memory_alignment(MemNode* s, int iv_adjust);
410 // (Start, end] half-open range defining which operands are vector
411 void vector_opd_range(Node* n, uint* start, uint* end);
412 // Smallest type containing range of values
413 const Type* container_type(Node* n);
414 // Adjust pre-loop limit so that in main loop, a load/store reference
415 // to align_to_ref will be a position zero in the vector.
416 void align_initial_loop_index(MemNode* align_to_ref);
417 // Find pre loop end from main loop. Returns null if none.
418 CountedLoopEndNode* get_pre_loop_end(CountedLoopNode *cl);
419 // Is the use of d1 in u1 at the same operand position as d2 in u2?
420 bool opnd_positions_match(Node* d1, Node* u1, Node* d2, Node* u2);
421 void init();
422
423 // print methods
424 void print_packset();
425 void print_pack(Node_List* p);
426 void print_bb();
427 void print_stmt(Node* s);
428 char* blank(uint depth);
429
430 void packset_sort(int n);
431 };
432
433
434
435 //------------------------------SWPointer---------------------------
436 // Information about an address for dependence checking and vector alignment
437 class SWPointer VALUE_OBJ_CLASS_SPEC {
438 protected:
439 MemNode* _mem; // My memory reference node
440 SuperWord* _slp; // SuperWord class
441
442 Node* _base; // NULL if unsafe nonheap reference
443 Node* _adr; // address pointer
444 jint _scale; // multipler for iv (in bytes), 0 if no loop iv
445 jint _offset; // constant offset (in bytes)
446 Node* _invar; // invariant offset (in bytes), NULL if none
447 bool _negate_invar; // if true then use: (0 - _invar)
448 Node_Stack* _nstack; // stack used to record a swpointer trace of variants
449 bool _analyze_only; // Used in loop unrolling only for swpointer trace
450 uint _stack_idx; // Used in loop unrolling only for swpointer trace
451
452 PhaseIdealLoop* phase() { return _slp->phase(); }
453 IdealLoopTree* lpt() { return _slp->lpt(); }
454 PhiNode* iv() { return _slp->iv(); } // Induction var
455
456 bool invariant(Node* n) {
457 Node *n_c = phase()->get_ctrl(n);
458 return !lpt()->is_member(phase()->get_loop(n_c));
459 }
460
461 // Match: k*iv + offset
462 bool scaled_iv_plus_offset(Node* n);
463 // Match: k*iv where k is a constant that's not zero
464 bool scaled_iv(Node* n);
465 // Match: offset is (k [+/- invariant])
466 bool offset_plus_k(Node* n, bool negate = false);
467
468 public:
469 enum CMP {
470 Less = 1,
471 Greater = 2,
472 Equal = 4,
473 NotEqual = (Less | Greater),
474 NotComparable = (Less | Greater | Equal)
475 };
476
477 SWPointer(MemNode* mem, SuperWord* slp, Node_Stack *nstack, bool analyze_only);
478 // Following is used to create a temporary object during
479 // the pattern match of an address expression.
480 SWPointer(SWPointer* p);
481
482 bool valid() { return _adr != NULL; }
483 bool has_iv() { return _scale != 0; }
484
485 Node* base() { return _base; }
486 Node* adr() { return _adr; }
487 MemNode* mem() { return _mem; }
488 int scale_in_bytes() { return _scale; }
489 Node* invar() { return _invar; }
490 bool negate_invar() { return _negate_invar; }
491 int offset_in_bytes() { return _offset; }
492 int memory_size() { return _mem->memory_size(); }
493 Node_Stack* node_stack() { return _nstack; }
494
495 // Comparable?
496 int cmp(SWPointer& q) {
497 if (valid() && q.valid() &&
498 (_adr == q._adr || _base == _adr && q._base == q._adr) &&
499 _scale == q._scale &&
500 _invar == q._invar &&
501 _negate_invar == q._negate_invar) {
502 bool overlap = q._offset < _offset + memory_size() &&
503 _offset < q._offset + q.memory_size();
504 return overlap ? Equal : (_offset < q._offset ? Less : Greater);
505 } else {
506 return NotComparable;
507 }
508 }
509
510 bool not_equal(SWPointer& q) { return not_equal(cmp(q)); }
511 bool equal(SWPointer& q) { return equal(cmp(q)); }
512 bool comparable(SWPointer& q) { return comparable(cmp(q)); }
513 static bool not_equal(int cmp) { return cmp <= NotEqual; }
514 static bool equal(int cmp) { return cmp == Equal; }
515 static bool comparable(int cmp) { return cmp < NotComparable; }
516
517 void print();
518 };
519
520
521 //------------------------------OrderedPair---------------------------
522 // Ordered pair of Node*.
523 class OrderedPair VALUE_OBJ_CLASS_SPEC {
524 protected:
525 Node* _p1;
526 Node* _p2;
527 public:
528 OrderedPair() : _p1(NULL), _p2(NULL) {}
529 OrderedPair(Node* p1, Node* p2) {
530 if (p1->_idx < p2->_idx) {
531 _p1 = p1; _p2 = p2;
532 } else {
533 _p1 = p2; _p2 = p1;
534 }
535 }
536
537 bool operator==(const OrderedPair &rhs) {
538 return _p1 == rhs._p1 && _p2 == rhs._p2;
539 }
540 void print() { tty->print(" (%d, %d)", _p1->_idx, _p2->_idx); }
541
542 static const OrderedPair initial;
543 };
544
545 #endif // SHARE_VM_OPTO_SUPERWORD_HPP