1 /*
2 * Copyright 2007 Sun Microsystems, Inc. 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 Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
20 * CA 95054 USA or visit www.sun.com if you need additional information or
21 * have any questions.
22 */
23
24 #include "incls/_precompiled.incl"
25 #include "incls/_superword.cpp.incl"
26
27 //
28 // S U P E R W O R D T R A N S F O R M
29 //=============================================================================
30
31 //------------------------------SuperWord---------------------------
32 SuperWord::SuperWord(PhaseIdealLoop* phase) :
33 _phase(phase),
34 _igvn(phase->_igvn),
35 _arena(phase->C->comp_arena()),
36 _packset(arena(), 8, 0, NULL), // packs for the current block
37 _bb_idx(arena(), (int)(1.10 * phase->C->unique()), 0, 0), // node idx to index in bb
38 _block(arena(), 8, 0, NULL), // nodes in current block
39 _data_entry(arena(), 8, 0, NULL), // nodes with all inputs from outside
40 _mem_slice_head(arena(), 8, 0, NULL), // memory slice heads
41 _mem_slice_tail(arena(), 8, 0, NULL), // memory slice tails
42 _node_info(arena(), 8, 0, SWNodeInfo::initial), // info needed per node
43 _align_to_ref(NULL), // memory reference to align vectors to
44 _disjoint_ptrs(arena(), 8, 0, OrderedPair::initial), // runtime disambiguated pointer pairs
45 _dg(_arena), // dependence graph
46 _visited(arena()), // visited node set
47 _post_visited(arena()), // post visited node set
48 _n_idx_list(arena(), 8), // scratch list of (node,index) pairs
49 _stk(arena(), 8, 0, NULL), // scratch stack of nodes
50 _nlist(arena(), 8, 0, NULL), // scratch list of nodes
51 _lpt(NULL), // loop tree node
52 _lp(NULL), // LoopNode
53 _bb(NULL), // basic block
54 _iv(NULL) // induction var
55 {}
56
57 //------------------------------transform_loop---------------------------
58 void SuperWord::transform_loop(IdealLoopTree* lpt) {
59 assert(lpt->_head->is_CountedLoop(), "must be");
60 CountedLoopNode *cl = lpt->_head->as_CountedLoop();
61
62 if (!cl->is_main_loop() ) return; // skip normal, pre, and post loops
63
64 // Check for no control flow in body (other than exit)
65 Node *cl_exit = cl->loopexit();
66 if (cl_exit->in(0) != lpt->_head) return;
67
68 // Check for pre-loop ending with CountedLoopEnd(Bool(Cmp(x,Opaque1(limit))))
69 CountedLoopEndNode* pre_end = get_pre_loop_end(cl);
70 if (pre_end == NULL) return;
71 Node *pre_opaq1 = pre_end->limit();
72 if (pre_opaq1->Opcode() != Op_Opaque1) return;
73
74 // Do vectors exist on this architecture?
75 if (vector_width_in_bytes() == 0) return;
76
77 init(); // initialize data structures
78
79 set_lpt(lpt);
80 set_lp(cl);
81
82 // For now, define one block which is the entire loop body
83 set_bb(cl);
84
85 assert(_packset.length() == 0, "packset must be empty");
86 SLP_extract();
87 }
88
89 //------------------------------SLP_extract---------------------------
90 // Extract the superword level parallelism
91 //
92 // 1) A reverse post-order of nodes in the block is constructed. By scanning
93 // this list from first to last, all definitions are visited before their uses.
94 //
95 // 2) A point-to-point dependence graph is constructed between memory references.
96 // This simplies the upcoming "independence" checker.
97 //
98 // 3) The maximum depth in the node graph from the beginning of the block
99 // to each node is computed. This is used to prune the graph search
100 // in the independence checker.
101 //
102 // 4) For integer types, the necessary bit width is propagated backwards
103 // from stores to allow packed operations on byte, char, and short
104 // integers. This reverses the promotion to type "int" that javac
105 // did for operations like: char c1,c2,c3; c1 = c2 + c3.
106 //
107 // 5) One of the memory references is picked to be an aligned vector reference.
108 // The pre-loop trip count is adjusted to align this reference in the
109 // unrolled body.
110 //
111 // 6) The initial set of pack pairs is seeded with memory references.
112 //
113 // 7) The set of pack pairs is extended by following use->def and def->use links.
114 //
115 // 8) The pairs are combined into vector sized packs.
116 //
117 // 9) Reorder the memory slices to co-locate members of the memory packs.
118 //
119 // 10) Generate ideal vector nodes for the final set of packs and where necessary,
120 // inserting scalar promotion, vector creation from multiple scalars, and
121 // extraction of scalar values from vectors.
122 //
123 void SuperWord::SLP_extract() {
124
125 // Ready the block
126
127 construct_bb();
128
129 dependence_graph();
130
131 compute_max_depth();
132
133 compute_vector_element_type();
134
135 // Attempt vectorization
136
137 find_adjacent_refs();
138
139 extend_packlist();
140
141 combine_packs();
142
143 construct_my_pack_map();
144
145 filter_packs();
146
147 schedule();
148
149 output();
150 }
151
152 //------------------------------find_adjacent_refs---------------------------
153 // Find the adjacent memory references and create pack pairs for them.
154 // This is the initial set of packs that will then be extended by
155 // following use->def and def->use links. The align positions are
156 // assigned relative to the reference "align_to_ref"
157 void SuperWord::find_adjacent_refs() {
158 // Get list of memory operations
159 Node_List memops;
160 for (int i = 0; i < _block.length(); i++) {
161 Node* n = _block.at(i);
162 if (n->is_Mem() && in_bb(n) &&
163 is_java_primitive(n->as_Mem()->memory_type())) {
164 int align = memory_alignment(n->as_Mem(), 0);
165 if (align != bottom_align) {
166 memops.push(n);
167 }
168 }
169 }
170 if (memops.size() == 0) return;
171
172 // Find a memory reference to align to. The pre-loop trip count
173 // is modified to align this reference to a vector-aligned address
174 find_align_to_ref(memops);
175 if (align_to_ref() == NULL) return;
176
177 SWPointer align_to_ref_p(align_to_ref(), this);
178 int offset = align_to_ref_p.offset_in_bytes();
179 int scale = align_to_ref_p.scale_in_bytes();
180 int vw = vector_width_in_bytes();
181 int stride_sign = (scale * iv_stride()) > 0 ? 1 : -1;
182 int iv_adjustment = (stride_sign * vw - (offset % vw)) % vw;
183
184 #ifndef PRODUCT
185 if (TraceSuperWord)
186 tty->print_cr("\noffset = %d iv_adjustment = %d elt_align = %d scale = %d iv_stride = %d",
187 offset, iv_adjustment, align_to_ref_p.memory_size(), align_to_ref_p.scale_in_bytes(), iv_stride());
188 #endif
189
190 // Set alignment relative to "align_to_ref"
191 for (int i = memops.size() - 1; i >= 0; i--) {
192 MemNode* s = memops.at(i)->as_Mem();
193 SWPointer p2(s, this);
194 if (p2.comparable(align_to_ref_p)) {
195 int align = memory_alignment(s, iv_adjustment);
196 set_alignment(s, align);
197 } else {
198 memops.remove(i);
199 }
200 }
201
202 // Create initial pack pairs of memory operations
203 for (uint i = 0; i < memops.size(); i++) {
204 Node* s1 = memops.at(i);
205 for (uint j = 0; j < memops.size(); j++) {
206 Node* s2 = memops.at(j);
207 if (s1 != s2 && are_adjacent_refs(s1, s2)) {
208 int align = alignment(s1);
209 if (stmts_can_pack(s1, s2, align)) {
210 Node_List* pair = new Node_List();
211 pair->push(s1);
212 pair->push(s2);
213 _packset.append(pair);
214 }
215 }
216 }
217 }
218
219 #ifndef PRODUCT
220 if (TraceSuperWord) {
221 tty->print_cr("\nAfter find_adjacent_refs");
222 print_packset();
223 }
224 #endif
225 }
226
227 //------------------------------find_align_to_ref---------------------------
228 // Find a memory reference to align the loop induction variable to.
229 // Looks first at stores then at loads, looking for a memory reference
230 // with the largest number of references similar to it.
231 void SuperWord::find_align_to_ref(Node_List &memops) {
232 GrowableArray<int> cmp_ct(arena(), memops.size(), memops.size(), 0);
233
234 // Count number of comparable memory ops
235 for (uint i = 0; i < memops.size(); i++) {
236 MemNode* s1 = memops.at(i)->as_Mem();
237 SWPointer p1(s1, this);
238 // Discard if pre loop can't align this reference
239 if (!ref_is_alignable(p1)) {
240 *cmp_ct.adr_at(i) = 0;
241 continue;
242 }
243 for (uint j = i+1; j < memops.size(); j++) {
244 MemNode* s2 = memops.at(j)->as_Mem();
245 if (isomorphic(s1, s2)) {
246 SWPointer p2(s2, this);
247 if (p1.comparable(p2)) {
248 (*cmp_ct.adr_at(i))++;
249 (*cmp_ct.adr_at(j))++;
250 }
251 }
252 }
253 }
254
255 // Find Store (or Load) with the greatest number of "comparable" references
256 int max_ct = 0;
257 int max_idx = -1;
258 int min_size = max_jint;
259 int min_iv_offset = max_jint;
260 for (uint j = 0; j < memops.size(); j++) {
261 MemNode* s = memops.at(j)->as_Mem();
262 if (s->is_Store()) {
263 SWPointer p(s, this);
264 if (cmp_ct.at(j) > max_ct ||
265 cmp_ct.at(j) == max_ct && (data_size(s) < min_size ||
266 data_size(s) == min_size &&
267 p.offset_in_bytes() < min_iv_offset)) {
268 max_ct = cmp_ct.at(j);
269 max_idx = j;
270 min_size = data_size(s);
271 min_iv_offset = p.offset_in_bytes();
272 }
273 }
274 }
275 // If no stores, look at loads
276 if (max_ct == 0) {
277 for (uint j = 0; j < memops.size(); j++) {
278 MemNode* s = memops.at(j)->as_Mem();
279 if (s->is_Load()) {
280 SWPointer p(s, this);
281 if (cmp_ct.at(j) > max_ct ||
282 cmp_ct.at(j) == max_ct && (data_size(s) < min_size ||
283 data_size(s) == min_size &&
284 p.offset_in_bytes() < min_iv_offset)) {
285 max_ct = cmp_ct.at(j);
286 max_idx = j;
287 min_size = data_size(s);
288 min_iv_offset = p.offset_in_bytes();
289 }
290 }
291 }
292 }
293
294 if (max_ct > 0)
295 set_align_to_ref(memops.at(max_idx)->as_Mem());
296
297 #ifndef PRODUCT
298 if (TraceSuperWord && Verbose) {
299 tty->print_cr("\nVector memops after find_align_to_refs");
300 for (uint i = 0; i < memops.size(); i++) {
301 MemNode* s = memops.at(i)->as_Mem();
302 s->dump();
303 }
304 }
305 #endif
306 }
307
308 //------------------------------ref_is_alignable---------------------------
309 // Can the preloop align the reference to position zero in the vector?
310 bool SuperWord::ref_is_alignable(SWPointer& p) {
311 if (!p.has_iv()) {
312 return true; // no induction variable
313 }
314 CountedLoopEndNode* pre_end = get_pre_loop_end(lp()->as_CountedLoop());
315 assert(pre_end->stride_is_con(), "pre loop stride is constant");
316 int preloop_stride = pre_end->stride_con();
317
318 int span = preloop_stride * p.scale_in_bytes();
319
320 // Stride one accesses are alignable.
321 if (ABS(span) == p.memory_size())
322 return true;
323
324 // If initial offset from start of object is computable,
325 // compute alignment within the vector.
326 int vw = vector_width_in_bytes();
327 if (vw % span == 0) {
328 Node* init_nd = pre_end->init_trip();
329 if (init_nd->is_Con() && p.invar() == NULL) {
330 int init = init_nd->bottom_type()->is_int()->get_con();
331
332 int init_offset = init * p.scale_in_bytes() + p.offset_in_bytes();
333 assert(init_offset >= 0, "positive offset from object start");
334
335 if (span > 0) {
336 return (vw - (init_offset % vw)) % span == 0;
337 } else {
338 assert(span < 0, "nonzero stride * scale");
339 return (init_offset % vw) % -span == 0;
340 }
341 }
342 }
343 return false;
344 }
345
346 //---------------------------dependence_graph---------------------------
347 // Construct dependency graph.
348 // Add dependence edges to load/store nodes for memory dependence
349 // A.out()->DependNode.in(1) and DependNode.out()->B.prec(x)
350 void SuperWord::dependence_graph() {
351 // First, assign a dependence node to each memory node
352 for (int i = 0; i < _block.length(); i++ ) {
353 Node *n = _block.at(i);
354 if (n->is_Mem() || n->is_Phi() && n->bottom_type() == Type::MEMORY) {
355 _dg.make_node(n);
356 }
357 }
358
359 // For each memory slice, create the dependences
360 for (int i = 0; i < _mem_slice_head.length(); i++) {
361 Node* n = _mem_slice_head.at(i);
362 Node* n_tail = _mem_slice_tail.at(i);
363
364 // Get slice in predecessor order (last is first)
365 mem_slice_preds(n_tail, n, _nlist);
366
367 // Make the slice dependent on the root
368 DepMem* slice = _dg.dep(n);
369 _dg.make_edge(_dg.root(), slice);
370
371 // Create a sink for the slice
372 DepMem* slice_sink = _dg.make_node(NULL);
373 _dg.make_edge(slice_sink, _dg.tail());
374
375 // Now visit each pair of memory ops, creating the edges
376 for (int j = _nlist.length() - 1; j >= 0 ; j--) {
377 Node* s1 = _nlist.at(j);
378
379 // If no dependency yet, use slice
380 if (_dg.dep(s1)->in_cnt() == 0) {
381 _dg.make_edge(slice, s1);
382 }
383 SWPointer p1(s1->as_Mem(), this);
384 bool sink_dependent = true;
385 for (int k = j - 1; k >= 0; k--) {
386 Node* s2 = _nlist.at(k);
387 if (s1->is_Load() && s2->is_Load())
388 continue;
389 SWPointer p2(s2->as_Mem(), this);
390
391 int cmp = p1.cmp(p2);
392 if (SuperWordRTDepCheck &&
393 p1.base() != p2.base() && p1.valid() && p2.valid()) {
394 // Create a runtime check to disambiguate
395 OrderedPair pp(p1.base(), p2.base());
396 _disjoint_ptrs.append_if_missing(pp);
397 } else if (!SWPointer::not_equal(cmp)) {
398 // Possibly same address
399 _dg.make_edge(s1, s2);
400 sink_dependent = false;
401 }
402 }
403 if (sink_dependent) {
404 _dg.make_edge(s1, slice_sink);
405 }
406 }
407 #ifndef PRODUCT
408 if (TraceSuperWord) {
409 tty->print_cr("\nDependence graph for slice: %d", n->_idx);
410 for (int q = 0; q < _nlist.length(); q++) {
411 _dg.print(_nlist.at(q));
412 }
413 tty->cr();
414 }
415 #endif
416 _nlist.clear();
417 }
418
419 #ifndef PRODUCT
420 if (TraceSuperWord) {
421 tty->print_cr("\ndisjoint_ptrs: %s", _disjoint_ptrs.length() > 0 ? "" : "NONE");
422 for (int r = 0; r < _disjoint_ptrs.length(); r++) {
423 _disjoint_ptrs.at(r).print();
424 tty->cr();
425 }
426 tty->cr();
427 }
428 #endif
429 }
430
431 //---------------------------mem_slice_preds---------------------------
432 // Return a memory slice (node list) in predecessor order starting at "start"
433 void SuperWord::mem_slice_preds(Node* start, Node* stop, GrowableArray<Node*> &preds) {
434 assert(preds.length() == 0, "start empty");
435 Node* n = start;
436 Node* prev = NULL;
437 while (true) {
438 assert(in_bb(n), "must be in block");
439 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
440 Node* out = n->fast_out(i);
441 if (out->is_Load()) {
442 if (in_bb(out)) {
443 preds.push(out);
444 }
445 } else {
446 // FIXME
447 if (out->is_MergeMem() && !in_bb(out)) {
448 // Either unrolling is causing a memory edge not to disappear,
449 // or need to run igvn.optimize() again before SLP
450 } else if (out->is_Phi() && out->bottom_type() == Type::MEMORY && !in_bb(out)) {
451 // Ditto. Not sure what else to check further.
452 } else if (out->Opcode() == Op_StoreCM && out->in(4) == n) {
453 // StoreCM has an input edge used as a precedence edge.
454 // Maybe an issue when oop stores are vectorized.
455 } else {
456 assert(out == prev || prev == NULL, "no branches off of store slice");
457 }
458 }
459 }
460 if (n == stop) break;
461 preds.push(n);
462 prev = n;
463 n = n->in(MemNode::Memory);
464 }
465 }
466
467 //------------------------------stmts_can_pack---------------------------
468 // Can s1 and s2 be in a pack with s1 immediately preceeding s2 and
469 // s1 aligned at "align"
470 bool SuperWord::stmts_can_pack(Node* s1, Node* s2, int align) {
471 if (isomorphic(s1, s2)) {
472 if (independent(s1, s2)) {
473 if (!exists_at(s1, 0) && !exists_at(s2, 1)) {
474 if (!s1->is_Mem() || are_adjacent_refs(s1, s2)) {
475 int s1_align = alignment(s1);
476 int s2_align = alignment(s2);
477 if (s1_align == top_align || s1_align == align) {
478 if (s2_align == top_align || s2_align == align + data_size(s1)) {
479 return true;
480 }
481 }
482 }
483 }
484 }
485 }
486 return false;
487 }
488
489 //------------------------------exists_at---------------------------
490 // Does s exist in a pack at position pos?
491 bool SuperWord::exists_at(Node* s, uint pos) {
492 for (int i = 0; i < _packset.length(); i++) {
493 Node_List* p = _packset.at(i);
494 if (p->at(pos) == s) {
495 return true;
496 }
497 }
498 return false;
499 }
500
501 //------------------------------are_adjacent_refs---------------------------
502 // Is s1 immediately before s2 in memory?
503 bool SuperWord::are_adjacent_refs(Node* s1, Node* s2) {
504 if (!s1->is_Mem() || !s2->is_Mem()) return false;
505 if (!in_bb(s1) || !in_bb(s2)) return false;
506 // FIXME - co_locate_pack fails on Stores in different mem-slices, so
507 // only pack memops that are in the same alias set until that's fixed.
508 if (_phase->C->get_alias_index(s1->as_Mem()->adr_type()) !=
509 _phase->C->get_alias_index(s2->as_Mem()->adr_type()))
510 return false;
511 SWPointer p1(s1->as_Mem(), this);
512 SWPointer p2(s2->as_Mem(), this);
513 if (p1.base() != p2.base() || !p1.comparable(p2)) return false;
514 int diff = p2.offset_in_bytes() - p1.offset_in_bytes();
515 return diff == data_size(s1);
516 }
517
518 //------------------------------isomorphic---------------------------
519 // Are s1 and s2 similar?
520 bool SuperWord::isomorphic(Node* s1, Node* s2) {
521 if (s1->Opcode() != s2->Opcode()) return false;
522 if (s1->req() != s2->req()) return false;
523 if (s1->in(0) != s2->in(0)) return false;
524 if (velt_type(s1) != velt_type(s2)) return false;
525 return true;
526 }
527
528 //------------------------------independent---------------------------
529 // Is there no data path from s1 to s2 or s2 to s1?
530 bool SuperWord::independent(Node* s1, Node* s2) {
531 // assert(s1->Opcode() == s2->Opcode(), "check isomorphic first");
532 int d1 = depth(s1);
533 int d2 = depth(s2);
534 if (d1 == d2) return s1 != s2;
535 Node* deep = d1 > d2 ? s1 : s2;
536 Node* shallow = d1 > d2 ? s2 : s1;
537
538 visited_clear();
539
540 return independent_path(shallow, deep);
541 }
542
543 //------------------------------independent_path------------------------------
544 // Helper for independent
545 bool SuperWord::independent_path(Node* shallow, Node* deep, uint dp) {
546 if (dp >= 1000) return false; // stop deep recursion
547 visited_set(deep);
548 int shal_depth = depth(shallow);
549 assert(shal_depth <= depth(deep), "must be");
550 for (DepPreds preds(deep, _dg); !preds.done(); preds.next()) {
551 Node* pred = preds.current();
552 if (in_bb(pred) && !visited_test(pred)) {
553 if (shallow == pred) {
554 return false;
555 }
556 if (shal_depth < depth(pred) && !independent_path(shallow, pred, dp+1)) {
557 return false;
558 }
559 }
560 }
561 return true;
562 }
563
564 //------------------------------set_alignment---------------------------
565 void SuperWord::set_alignment(Node* s1, Node* s2, int align) {
566 set_alignment(s1, align);
567 set_alignment(s2, align + data_size(s1));
568 }
569
570 //------------------------------data_size---------------------------
571 int SuperWord::data_size(Node* s) {
572 const Type* t = velt_type(s);
573 BasicType bt = t->array_element_basic_type();
574 int bsize = type2aelembytes(bt);
575 assert(bsize != 0, "valid size");
576 return bsize;
577 }
578
579 //------------------------------extend_packlist---------------------------
580 // Extend packset by following use->def and def->use links from pack members.
581 void SuperWord::extend_packlist() {
582 bool changed;
583 do {
584 changed = false;
585 for (int i = 0; i < _packset.length(); i++) {
586 Node_List* p = _packset.at(i);
587 changed |= follow_use_defs(p);
588 changed |= follow_def_uses(p);
589 }
590 } while (changed);
591
592 #ifndef PRODUCT
593 if (TraceSuperWord) {
594 tty->print_cr("\nAfter extend_packlist");
595 print_packset();
596 }
597 #endif
598 }
599
600 //------------------------------follow_use_defs---------------------------
601 // Extend the packset by visiting operand definitions of nodes in pack p
602 bool SuperWord::follow_use_defs(Node_List* p) {
603 Node* s1 = p->at(0);
604 Node* s2 = p->at(1);
605 assert(p->size() == 2, "just checking");
606 assert(s1->req() == s2->req(), "just checking");
607 assert(alignment(s1) + data_size(s1) == alignment(s2), "just checking");
608
609 if (s1->is_Load()) return false;
610
611 int align = alignment(s1);
612 bool changed = false;
613 int start = s1->is_Store() ? MemNode::ValueIn : 1;
614 int end = s1->is_Store() ? MemNode::ValueIn+1 : s1->req();
615 for (int j = start; j < end; j++) {
616 Node* t1 = s1->in(j);
617 Node* t2 = s2->in(j);
618 if (!in_bb(t1) || !in_bb(t2))
619 continue;
620 if (stmts_can_pack(t1, t2, align)) {
621 if (est_savings(t1, t2) >= 0) {
622 Node_List* pair = new Node_List();
623 pair->push(t1);
624 pair->push(t2);
625 _packset.append(pair);
626 set_alignment(t1, t2, align);
627 changed = true;
628 }
629 }
630 }
631 return changed;
632 }
633
634 //------------------------------follow_def_uses---------------------------
635 // Extend the packset by visiting uses of nodes in pack p
636 bool SuperWord::follow_def_uses(Node_List* p) {
637 bool changed = false;
638 Node* s1 = p->at(0);
639 Node* s2 = p->at(1);
640 assert(p->size() == 2, "just checking");
641 assert(s1->req() == s2->req(), "just checking");
642 assert(alignment(s1) + data_size(s1) == alignment(s2), "just checking");
643
644 if (s1->is_Store()) return false;
645
646 int align = alignment(s1);
647 int savings = -1;
648 Node* u1 = NULL;
649 Node* u2 = NULL;
650 for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
651 Node* t1 = s1->fast_out(i);
652 if (!in_bb(t1)) continue;
653 for (DUIterator_Fast jmax, j = s2->fast_outs(jmax); j < jmax; j++) {
654 Node* t2 = s2->fast_out(j);
655 if (!in_bb(t2)) continue;
656 if (!opnd_positions_match(s1, t1, s2, t2))
657 continue;
658 if (stmts_can_pack(t1, t2, align)) {
659 int my_savings = est_savings(t1, t2);
660 if (my_savings > savings) {
661 savings = my_savings;
662 u1 = t1;
663 u2 = t2;
664 }
665 }
666 }
667 }
668 if (savings >= 0) {
669 Node_List* pair = new Node_List();
670 pair->push(u1);
671 pair->push(u2);
672 _packset.append(pair);
673 set_alignment(u1, u2, align);
674 changed = true;
675 }
676 return changed;
677 }
678
679 //---------------------------opnd_positions_match-------------------------
680 // Is the use of d1 in u1 at the same operand position as d2 in u2?
681 bool SuperWord::opnd_positions_match(Node* d1, Node* u1, Node* d2, Node* u2) {
682 uint ct = u1->req();
683 if (ct != u2->req()) return false;
684 uint i1 = 0;
685 uint i2 = 0;
686 do {
687 for (i1++; i1 < ct; i1++) if (u1->in(i1) == d1) break;
688 for (i2++; i2 < ct; i2++) if (u2->in(i2) == d2) break;
689 if (i1 != i2) {
690 return false;
691 }
692 } while (i1 < ct);
693 return true;
694 }
695
696 //------------------------------est_savings---------------------------
697 // Estimate the savings from executing s1 and s2 as a pack
698 int SuperWord::est_savings(Node* s1, Node* s2) {
699 int save = 2 - 1; // 2 operations per instruction in packed form
700
701 // inputs
702 for (uint i = 1; i < s1->req(); i++) {
703 Node* x1 = s1->in(i);
704 Node* x2 = s2->in(i);
705 if (x1 != x2) {
706 if (are_adjacent_refs(x1, x2)) {
707 save += adjacent_profit(x1, x2);
708 } else if (!in_packset(x1, x2)) {
709 save -= pack_cost(2);
710 } else {
711 save += unpack_cost(2);
712 }
713 }
714 }
715
716 // uses of result
717 uint ct = 0;
718 for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
719 Node* s1_use = s1->fast_out(i);
720 for (int j = 0; j < _packset.length(); j++) {
721 Node_List* p = _packset.at(j);
722 if (p->at(0) == s1_use) {
723 for (DUIterator_Fast kmax, k = s2->fast_outs(kmax); k < kmax; k++) {
724 Node* s2_use = s2->fast_out(k);
725 if (p->at(p->size()-1) == s2_use) {
726 ct++;
727 if (are_adjacent_refs(s1_use, s2_use)) {
728 save += adjacent_profit(s1_use, s2_use);
729 }
730 }
731 }
732 }
733 }
734 }
735
736 if (ct < s1->outcnt()) save += unpack_cost(1);
737 if (ct < s2->outcnt()) save += unpack_cost(1);
738
739 return save;
740 }
741
742 //------------------------------costs---------------------------
743 int SuperWord::adjacent_profit(Node* s1, Node* s2) { return 2; }
744 int SuperWord::pack_cost(int ct) { return ct; }
745 int SuperWord::unpack_cost(int ct) { return ct; }
746
747 //------------------------------combine_packs---------------------------
748 // Combine packs A and B with A.last == B.first into A.first..,A.last,B.second,..B.last
749 void SuperWord::combine_packs() {
750 bool changed;
751 do {
752 changed = false;
753 for (int i = 0; i < _packset.length(); i++) {
754 Node_List* p1 = _packset.at(i);
755 if (p1 == NULL) continue;
756 for (int j = 0; j < _packset.length(); j++) {
757 Node_List* p2 = _packset.at(j);
758 if (p2 == NULL) continue;
759 if (p1->at(p1->size()-1) == p2->at(0)) {
760 for (uint k = 1; k < p2->size(); k++) {
761 p1->push(p2->at(k));
762 }
763 _packset.at_put(j, NULL);
764 changed = true;
765 }
766 }
767 }
768 } while (changed);
769
770 for (int i = _packset.length() - 1; i >= 0; i--) {
771 Node_List* p1 = _packset.at(i);
772 if (p1 == NULL) {
773 _packset.remove_at(i);
774 }
775 }
776
777 #ifndef PRODUCT
778 if (TraceSuperWord) {
779 tty->print_cr("\nAfter combine_packs");
780 print_packset();
781 }
782 #endif
783 }
784
785 //-----------------------------construct_my_pack_map--------------------------
786 // Construct the map from nodes to packs. Only valid after the
787 // point where a node is only in one pack (after combine_packs).
788 void SuperWord::construct_my_pack_map() {
789 Node_List* rslt = NULL;
790 for (int i = 0; i < _packset.length(); i++) {
791 Node_List* p = _packset.at(i);
792 for (uint j = 0; j < p->size(); j++) {
793 Node* s = p->at(j);
794 assert(my_pack(s) == NULL, "only in one pack");
795 set_my_pack(s, p);
796 }
797 }
798 }
799
800 //------------------------------filter_packs---------------------------
801 // Remove packs that are not implemented or not profitable.
802 void SuperWord::filter_packs() {
803
804 // Remove packs that are not implemented
805 for (int i = _packset.length() - 1; i >= 0; i--) {
806 Node_List* pk = _packset.at(i);
807 bool impl = implemented(pk);
808 if (!impl) {
809 #ifndef PRODUCT
810 if (TraceSuperWord && Verbose) {
811 tty->print_cr("Unimplemented");
812 pk->at(0)->dump();
813 }
814 #endif
815 remove_pack_at(i);
816 }
817 }
818
819 // Remove packs that are not profitable
820 bool changed;
821 do {
822 changed = false;
823 for (int i = _packset.length() - 1; i >= 0; i--) {
824 Node_List* pk = _packset.at(i);
825 bool prof = profitable(pk);
826 if (!prof) {
827 #ifndef PRODUCT
828 if (TraceSuperWord && Verbose) {
829 tty->print_cr("Unprofitable");
830 pk->at(0)->dump();
831 }
832 #endif
833 remove_pack_at(i);
834 changed = true;
835 }
836 }
837 } while (changed);
838
839 #ifndef PRODUCT
840 if (TraceSuperWord) {
841 tty->print_cr("\nAfter filter_packs");
842 print_packset();
843 tty->cr();
844 }
845 #endif
846 }
847
848 //------------------------------implemented---------------------------
849 // Can code be generated for pack p?
850 bool SuperWord::implemented(Node_List* p) {
851 Node* p0 = p->at(0);
852 int vopc = VectorNode::opcode(p0->Opcode(), p->size(), velt_type(p0));
853 return vopc > 0 && Matcher::has_match_rule(vopc);
854 }
855
856 //------------------------------profitable---------------------------
857 // For pack p, are all operands and all uses (with in the block) vector?
858 bool SuperWord::profitable(Node_List* p) {
859 Node* p0 = p->at(0);
860 uint start, end;
861 vector_opd_range(p0, &start, &end);
862
863 // Return false if some input is not vector and inside block
864 for (uint i = start; i < end; i++) {
865 if (!is_vector_use(p0, i)) {
866 // For now, return false if not scalar promotion case (inputs are the same.)
867 // Later, implement PackNode and allow differring, non-vector inputs
868 // (maybe just the ones from outside the block.)
869 Node* p0_def = p0->in(i);
870 for (uint j = 1; j < p->size(); j++) {
871 Node* use = p->at(j);
872 Node* def = use->in(i);
873 if (p0_def != def)
874 return false;
875 }
876 }
877 }
878 if (!p0->is_Store()) {
879 // For now, return false if not all uses are vector.
880 // Later, implement ExtractNode and allow non-vector uses (maybe
881 // just the ones outside the block.)
882 for (uint i = 0; i < p->size(); i++) {
883 Node* def = p->at(i);
884 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
885 Node* use = def->fast_out(j);
886 for (uint k = 0; k < use->req(); k++) {
887 Node* n = use->in(k);
888 if (def == n) {
889 if (!is_vector_use(use, k)) {
890 return false;
891 }
892 }
893 }
894 }
895 }
896 }
897 return true;
898 }
899
900 //------------------------------schedule---------------------------
901 // Adjust the memory graph for the packed operations
902 void SuperWord::schedule() {
903
904 // Co-locate in the memory graph the members of each memory pack
905 for (int i = 0; i < _packset.length(); i++) {
906 co_locate_pack(_packset.at(i));
907 }
908 }
909
910 //------------------------------co_locate_pack---------------------------
911 // Within a pack, move stores down to the last executed store,
912 // and move loads up to the first executed load.
913 void SuperWord::co_locate_pack(Node_List* pk) {
914 if (pk->at(0)->is_Store()) {
915 // Push Stores down towards last executed pack member
916 MemNode* first = executed_first(pk)->as_Mem();
917 MemNode* last = executed_last(pk)->as_Mem();
918 MemNode* insert_pt = last;
919 MemNode* current = last->in(MemNode::Memory)->as_Mem();
920 while (true) {
921 assert(in_bb(current), "stay in block");
922 Node* my_mem = current->in(MemNode::Memory);
923 if (in_pack(current, pk)) {
924 // Forward users of my memory state to my input memory state
925 _igvn.hash_delete(current);
926 _igvn.hash_delete(my_mem);
927 for (DUIterator i = current->outs(); current->has_out(i); i++) {
928 Node* use = current->out(i);
929 if (use->is_Mem()) {
930 assert(use->in(MemNode::Memory) == current, "must be");
931 _igvn.hash_delete(use);
932 use->set_req(MemNode::Memory, my_mem);
933 _igvn._worklist.push(use);
934 --i; // deleted this edge; rescan position
935 }
936 }
937 // put current immediately before insert_pt
938 current->set_req(MemNode::Memory, insert_pt->in(MemNode::Memory));
939 _igvn.hash_delete(insert_pt);
940 insert_pt->set_req(MemNode::Memory, current);
941 _igvn._worklist.push(insert_pt);
942 _igvn._worklist.push(current);
943 insert_pt = current;
944 }
945 if (current == first) break;
946 current = my_mem->as_Mem();
947 }
948 } else if (pk->at(0)->is_Load()) {
949 // Pull Loads up towards first executed pack member
950 LoadNode* first = executed_first(pk)->as_Load();
951 Node* first_mem = first->in(MemNode::Memory);
952 _igvn.hash_delete(first_mem);
953 // Give each load same memory state as first
954 for (uint i = 0; i < pk->size(); i++) {
955 LoadNode* ld = pk->at(i)->as_Load();
956 _igvn.hash_delete(ld);
957 ld->set_req(MemNode::Memory, first_mem);
958 _igvn._worklist.push(ld);
959 }
960 }
961 }
962
963 //------------------------------output---------------------------
964 // Convert packs into vector node operations
965 void SuperWord::output() {
966 if (_packset.length() == 0) return;
967
968 // MUST ENSURE main loop's initial value is properly aligned:
969 // (iv_initial_value + min_iv_offset) % vector_width_in_bytes() == 0
970
971 align_initial_loop_index(align_to_ref());
972
973 // Insert extract (unpack) operations for scalar uses
974 for (int i = 0; i < _packset.length(); i++) {
975 insert_extracts(_packset.at(i));
976 }
977
978 for (int i = 0; i < _block.length(); i++) {
979 Node* n = _block.at(i);
980 Node_List* p = my_pack(n);
981 if (p && n == executed_last(p)) {
982 uint vlen = p->size();
983 Node* vn = NULL;
984 Node* low_adr = p->at(0);
985 Node* first = executed_first(p);
986 if (n->is_Load()) {
987 int opc = n->Opcode();
988 Node* ctl = n->in(MemNode::Control);
989 Node* mem = first->in(MemNode::Memory);
990 Node* adr = low_adr->in(MemNode::Address);
991 const TypePtr* atyp = n->adr_type();
992 vn = VectorLoadNode::make(_phase->C, opc, ctl, mem, adr, atyp, vlen);
993
994 } else if (n->is_Store()) {
995 // Promote value to be stored to vector
996 VectorNode* val = vector_opd(p, MemNode::ValueIn);
997
998 int opc = n->Opcode();
999 Node* ctl = n->in(MemNode::Control);
1000 Node* mem = first->in(MemNode::Memory);
1001 Node* adr = low_adr->in(MemNode::Address);
1002 const TypePtr* atyp = n->adr_type();
1003 vn = VectorStoreNode::make(_phase->C, opc, ctl, mem, adr, atyp, val, vlen);
1004
1005 } else if (n->req() == 3) {
1006 // Promote operands to vector
1007 Node* in1 = vector_opd(p, 1);
1008 Node* in2 = vector_opd(p, 2);
1009 vn = VectorNode::make(_phase->C, n->Opcode(), in1, in2, vlen, velt_type(n));
1010
1011 } else {
1012 ShouldNotReachHere();
1013 }
1014
1015 _phase->_igvn.register_new_node_with_optimizer(vn);
1016 _phase->set_ctrl(vn, _phase->get_ctrl(p->at(0)));
1017 for (uint j = 0; j < p->size(); j++) {
1018 Node* pm = p->at(j);
1019 _igvn.hash_delete(pm);
1020 _igvn.subsume_node(pm, vn);
1021 }
1022 _igvn._worklist.push(vn);
1023 }
1024 }
1025 }
1026
1027 //------------------------------vector_opd---------------------------
1028 // Create a vector operand for the nodes in pack p for operand: in(opd_idx)
1029 VectorNode* SuperWord::vector_opd(Node_List* p, int opd_idx) {
1030 Node* p0 = p->at(0);
1031 uint vlen = p->size();
1032 Node* opd = p0->in(opd_idx);
1033
1034 bool same_opd = true;
1035 for (uint i = 1; i < vlen; i++) {
1036 Node* pi = p->at(i);
1037 Node* in = pi->in(opd_idx);
1038 if (opd != in) {
1039 same_opd = false;
1040 break;
1041 }
1042 }
1043
1044 if (same_opd) {
1045 if (opd->is_Vector()) {
1046 return (VectorNode*)opd; // input is matching vector
1047 }
1048 // Convert scalar input to vector. Use p0's type because it's container
1049 // maybe smaller than the operand's container.
1050 const Type* opd_t = velt_type(!in_bb(opd) ? p0 : opd);
1051 const Type* p0_t = velt_type(p0);
1052 if (p0_t->higher_equal(opd_t)) opd_t = p0_t;
1053 VectorNode* vn = VectorNode::scalar2vector(_phase->C, opd, vlen, opd_t);
1054
1055 _phase->_igvn.register_new_node_with_optimizer(vn);
1056 _phase->set_ctrl(vn, _phase->get_ctrl(opd));
1057 return vn;
1058 }
1059
1060 // Insert pack operation
1061 const Type* opd_t = velt_type(!in_bb(opd) ? p0 : opd);
1062 PackNode* pk = PackNode::make(_phase->C, opd, opd_t);
1063
1064 for (uint i = 1; i < vlen; i++) {
1065 Node* pi = p->at(i);
1066 Node* in = pi->in(opd_idx);
1067 assert(my_pack(in) == NULL, "Should already have been unpacked");
1068 assert(opd_t == velt_type(!in_bb(in) ? pi : in), "all same type");
1069 pk->add_opd(in);
1070 }
1071 _phase->_igvn.register_new_node_with_optimizer(pk);
1072 _phase->set_ctrl(pk, _phase->get_ctrl(opd));
1073 return pk;
1074 }
1075
1076 //------------------------------insert_extracts---------------------------
1077 // If a use of pack p is not a vector use, then replace the
1078 // use with an extract operation.
1079 void SuperWord::insert_extracts(Node_List* p) {
1080 if (p->at(0)->is_Store()) return;
1081 assert(_n_idx_list.is_empty(), "empty (node,index) list");
1082
1083 // Inspect each use of each pack member. For each use that is
1084 // not a vector use, replace the use with an extract operation.
1085
1086 for (uint i = 0; i < p->size(); i++) {
1087 Node* def = p->at(i);
1088 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
1089 Node* use = def->fast_out(j);
1090 for (uint k = 0; k < use->req(); k++) {
1091 Node* n = use->in(k);
1092 if (def == n) {
1093 if (!is_vector_use(use, k)) {
1094 _n_idx_list.push(use, k);
1095 }
1096 }
1097 }
1098 }
1099 }
1100
1101 while (_n_idx_list.is_nonempty()) {
1102 Node* use = _n_idx_list.node();
1103 int idx = _n_idx_list.index();
1104 _n_idx_list.pop();
1105 Node* def = use->in(idx);
1106
1107 // Insert extract operation
1108 _igvn.hash_delete(def);
1109 _igvn.hash_delete(use);
1110 int def_pos = alignment(def) / data_size(def);
1111 const Type* def_t = velt_type(def);
1112
1113 Node* ex = ExtractNode::make(_phase->C, def, def_pos, def_t);
1114 _phase->_igvn.register_new_node_with_optimizer(ex);
1115 _phase->set_ctrl(ex, _phase->get_ctrl(def));
1116 use->set_req(idx, ex);
1117 _igvn._worklist.push(def);
1118 _igvn._worklist.push(use);
1119
1120 bb_insert_after(ex, bb_idx(def));
1121 set_velt_type(ex, def_t);
1122 }
1123 }
1124
1125 //------------------------------is_vector_use---------------------------
1126 // Is use->in(u_idx) a vector use?
1127 bool SuperWord::is_vector_use(Node* use, int u_idx) {
1128 Node_List* u_pk = my_pack(use);
1129 if (u_pk == NULL) return false;
1130 Node* def = use->in(u_idx);
1131 Node_List* d_pk = my_pack(def);
1132 if (d_pk == NULL) {
1133 // check for scalar promotion
1134 Node* n = u_pk->at(0)->in(u_idx);
1135 for (uint i = 1; i < u_pk->size(); i++) {
1136 if (u_pk->at(i)->in(u_idx) != n) return false;
1137 }
1138 return true;
1139 }
1140 if (u_pk->size() != d_pk->size())
1141 return false;
1142 for (uint i = 0; i < u_pk->size(); i++) {
1143 Node* ui = u_pk->at(i);
1144 Node* di = d_pk->at(i);
1145 if (ui->in(u_idx) != di || alignment(ui) != alignment(di))
1146 return false;
1147 }
1148 return true;
1149 }
1150
1151 //------------------------------construct_bb---------------------------
1152 // Construct reverse postorder list of block members
1153 void SuperWord::construct_bb() {
1154 Node* entry = bb();
1155
1156 assert(_stk.length() == 0, "stk is empty");
1157 assert(_block.length() == 0, "block is empty");
1158 assert(_data_entry.length() == 0, "data_entry is empty");
1159 assert(_mem_slice_head.length() == 0, "mem_slice_head is empty");
1160 assert(_mem_slice_tail.length() == 0, "mem_slice_tail is empty");
1161
1162 // Find non-control nodes with no inputs from within block,
1163 // create a temporary map from node _idx to bb_idx for use
1164 // by the visited and post_visited sets,
1165 // and count number of nodes in block.
1166 int bb_ct = 0;
1167 for (uint i = 0; i < lpt()->_body.size(); i++ ) {
1168 Node *n = lpt()->_body.at(i);
1169 set_bb_idx(n, i); // Create a temporary map
1170 if (in_bb(n)) {
1171 bb_ct++;
1172 if (!n->is_CFG()) {
1173 bool found = false;
1174 for (uint j = 0; j < n->req(); j++) {
1175 Node* def = n->in(j);
1176 if (def && in_bb(def)) {
1177 found = true;
1178 break;
1179 }
1180 }
1181 if (!found) {
1182 assert(n != entry, "can't be entry");
1183 _data_entry.push(n);
1184 }
1185 }
1186 }
1187 }
1188
1189 // Find memory slices (head and tail)
1190 for (DUIterator_Fast imax, i = lp()->fast_outs(imax); i < imax; i++) {
1191 Node *n = lp()->fast_out(i);
1192 if (in_bb(n) && (n->is_Phi() && n->bottom_type() == Type::MEMORY)) {
1193 Node* n_tail = n->in(LoopNode::LoopBackControl);
1194 _mem_slice_head.push(n);
1195 _mem_slice_tail.push(n_tail);
1196 }
1197 }
1198
1199 // Create an RPO list of nodes in block
1200
1201 visited_clear();
1202 post_visited_clear();
1203
1204 // Push all non-control nodes with no inputs from within block, then control entry
1205 for (int j = 0; j < _data_entry.length(); j++) {
1206 Node* n = _data_entry.at(j);
1207 visited_set(n);
1208 _stk.push(n);
1209 }
1210 visited_set(entry);
1211 _stk.push(entry);
1212
1213 // Do a depth first walk over out edges
1214 int rpo_idx = bb_ct - 1;
1215 int size;
1216 while ((size = _stk.length()) > 0) {
1217 Node* n = _stk.top(); // Leave node on stack
1218 if (!visited_test_set(n)) {
1219 // forward arc in graph
1220 } else if (!post_visited_test(n)) {
1221 // cross or back arc
1222 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1223 Node *use = n->fast_out(i);
1224 if (in_bb(use) && !visited_test(use) &&
1225 // Don't go around backedge
1226 (!use->is_Phi() || n == entry)) {
1227 _stk.push(use);
1228 }
1229 }
1230 if (_stk.length() == size) {
1231 // There were no additional uses, post visit node now
1232 _stk.pop(); // Remove node from stack
1233 assert(rpo_idx >= 0, "");
1234 _block.at_put_grow(rpo_idx, n);
1235 rpo_idx--;
1236 post_visited_set(n);
1237 assert(rpo_idx >= 0 || _stk.is_empty(), "");
1238 }
1239 } else {
1240 _stk.pop(); // Remove post-visited node from stack
1241 }
1242 }
1243
1244 // Create real map of block indices for nodes
1245 for (int j = 0; j < _block.length(); j++) {
1246 Node* n = _block.at(j);
1247 set_bb_idx(n, j);
1248 }
1249
1250 initialize_bb(); // Ensure extra info is allocated.
1251
1252 #ifndef PRODUCT
1253 if (TraceSuperWord) {
1254 print_bb();
1255 tty->print_cr("\ndata entry nodes: %s", _data_entry.length() > 0 ? "" : "NONE");
1256 for (int m = 0; m < _data_entry.length(); m++) {
1257 tty->print("%3d ", m);
1258 _data_entry.at(m)->dump();
1259 }
1260 tty->print_cr("\nmemory slices: %s", _mem_slice_head.length() > 0 ? "" : "NONE");
1261 for (int m = 0; m < _mem_slice_head.length(); m++) {
1262 tty->print("%3d ", m); _mem_slice_head.at(m)->dump();
1263 tty->print(" "); _mem_slice_tail.at(m)->dump();
1264 }
1265 }
1266 #endif
1267 assert(rpo_idx == -1 && bb_ct == _block.length(), "all block members found");
1268 }
1269
1270 //------------------------------initialize_bb---------------------------
1271 // Initialize per node info
1272 void SuperWord::initialize_bb() {
1273 Node* last = _block.at(_block.length() - 1);
1274 grow_node_info(bb_idx(last));
1275 }
1276
1277 //------------------------------bb_insert_after---------------------------
1278 // Insert n into block after pos
1279 void SuperWord::bb_insert_after(Node* n, int pos) {
1280 int n_pos = pos + 1;
1281 // Make room
1282 for (int i = _block.length() - 1; i >= n_pos; i--) {
1283 _block.at_put_grow(i+1, _block.at(i));
1284 }
1285 for (int j = _node_info.length() - 1; j >= n_pos; j--) {
1286 _node_info.at_put_grow(j+1, _node_info.at(j));
1287 }
1288 // Set value
1289 _block.at_put_grow(n_pos, n);
1290 _node_info.at_put_grow(n_pos, SWNodeInfo::initial);
1291 // Adjust map from node->_idx to _block index
1292 for (int i = n_pos; i < _block.length(); i++) {
1293 set_bb_idx(_block.at(i), i);
1294 }
1295 }
1296
1297 //------------------------------compute_max_depth---------------------------
1298 // Compute max depth for expressions from beginning of block
1299 // Use to prune search paths during test for independence.
1300 void SuperWord::compute_max_depth() {
1301 int ct = 0;
1302 bool again;
1303 do {
1304 again = false;
1305 for (int i = 0; i < _block.length(); i++) {
1306 Node* n = _block.at(i);
1307 if (!n->is_Phi()) {
1308 int d_orig = depth(n);
1309 int d_in = 0;
1310 for (DepPreds preds(n, _dg); !preds.done(); preds.next()) {
1311 Node* pred = preds.current();
1312 if (in_bb(pred)) {
1313 d_in = MAX2(d_in, depth(pred));
1314 }
1315 }
1316 if (d_in + 1 != d_orig) {
1317 set_depth(n, d_in + 1);
1318 again = true;
1319 }
1320 }
1321 }
1322 ct++;
1323 } while (again);
1324 #ifndef PRODUCT
1325 if (TraceSuperWord && Verbose)
1326 tty->print_cr("compute_max_depth iterated: %d times", ct);
1327 #endif
1328 }
1329
1330 //-------------------------compute_vector_element_type-----------------------
1331 // Compute necessary vector element type for expressions
1332 // This propagates backwards a narrower integer type when the
1333 // upper bits of the value are not needed.
1334 // Example: char a,b,c; a = b + c;
1335 // Normally the type of the add is integer, but for packed character
1336 // operations the type of the add needs to be char.
1337 void SuperWord::compute_vector_element_type() {
1338 #ifndef PRODUCT
1339 if (TraceSuperWord && Verbose)
1340 tty->print_cr("\ncompute_velt_type:");
1341 #endif
1342
1343 // Initial type
1344 for (int i = 0; i < _block.length(); i++) {
1345 Node* n = _block.at(i);
1346 const Type* t = n->is_Mem() ? Type::get_const_basic_type(n->as_Mem()->memory_type())
1347 : _igvn.type(n);
1348 const Type* vt = container_type(t);
1349 set_velt_type(n, vt);
1350 }
1351
1352 // Propagate narrowed type backwards through operations
1353 // that don't depend on higher order bits
1354 for (int i = _block.length() - 1; i >= 0; i--) {
1355 Node* n = _block.at(i);
1356 // Only integer types need be examined
1357 if (n->bottom_type()->isa_int()) {
1358 uint start, end;
1359 vector_opd_range(n, &start, &end);
1360 const Type* vt = velt_type(n);
1361
1362 for (uint j = start; j < end; j++) {
1363 Node* in = n->in(j);
1364 // Don't propagate through a type conversion
1365 if (n->bottom_type() != in->bottom_type())
1366 continue;
1367 switch(in->Opcode()) {
1368 case Op_AddI: case Op_AddL:
1369 case Op_SubI: case Op_SubL:
1370 case Op_MulI: case Op_MulL:
1371 case Op_AndI: case Op_AndL:
1372 case Op_OrI: case Op_OrL:
1373 case Op_XorI: case Op_XorL:
1374 case Op_LShiftI: case Op_LShiftL:
1375 case Op_CMoveI: case Op_CMoveL:
1376 if (in_bb(in)) {
1377 bool same_type = true;
1378 for (DUIterator_Fast kmax, k = in->fast_outs(kmax); k < kmax; k++) {
1379 Node *use = in->fast_out(k);
1380 if (!in_bb(use) || velt_type(use) != vt) {
1381 same_type = false;
1382 break;
1383 }
1384 }
1385 if (same_type) {
1386 set_velt_type(in, vt);
1387 }
1388 }
1389 }
1390 }
1391 }
1392 }
1393 #ifndef PRODUCT
1394 if (TraceSuperWord && Verbose) {
1395 for (int i = 0; i < _block.length(); i++) {
1396 Node* n = _block.at(i);
1397 velt_type(n)->dump();
1398 tty->print("\t");
1399 n->dump();
1400 }
1401 }
1402 #endif
1403 }
1404
1405 //------------------------------memory_alignment---------------------------
1406 // Alignment within a vector memory reference
1407 int SuperWord::memory_alignment(MemNode* s, int iv_adjust_in_bytes) {
1408 SWPointer p(s, this);
1409 if (!p.valid()) {
1410 return bottom_align;
1411 }
1412 int offset = p.offset_in_bytes();
1413 offset += iv_adjust_in_bytes;
1414 int off_rem = offset % vector_width_in_bytes();
1415 int off_mod = off_rem >= 0 ? off_rem : off_rem + vector_width_in_bytes();
1416 return off_mod;
1417 }
1418
1419 //---------------------------container_type---------------------------
1420 // Smallest type containing range of values
1421 const Type* SuperWord::container_type(const Type* t) {
1422 if (t->isa_aryptr()) {
1423 t = t->is_aryptr()->elem();
1424 }
1425 if (t->basic_type() == T_INT) {
1426 if (t->higher_equal(TypeInt::BOOL)) return TypeInt::BOOL;
1427 if (t->higher_equal(TypeInt::BYTE)) return TypeInt::BYTE;
1428 if (t->higher_equal(TypeInt::CHAR)) return TypeInt::CHAR;
1429 if (t->higher_equal(TypeInt::SHORT)) return TypeInt::SHORT;
1430 return TypeInt::INT;
1431 }
1432 return t;
1433 }
1434
1435 //-------------------------vector_opd_range-----------------------
1436 // (Start, end] half-open range defining which operands are vector
1437 void SuperWord::vector_opd_range(Node* n, uint* start, uint* end) {
1438 switch (n->Opcode()) {
1439 case Op_LoadB: case Op_LoadC:
1440 case Op_LoadI: case Op_LoadL:
1441 case Op_LoadF: case Op_LoadD:
1442 case Op_LoadP:
1443 *start = 0;
1444 *end = 0;
1445 return;
1446 case Op_StoreB: case Op_StoreC:
1447 case Op_StoreI: case Op_StoreL:
1448 case Op_StoreF: case Op_StoreD:
1449 case Op_StoreP:
1450 *start = MemNode::ValueIn;
1451 *end = *start + 1;
1452 return;
1453 case Op_LShiftI: case Op_LShiftL:
1454 *start = 1;
1455 *end = 2;
1456 return;
1457 case Op_CMoveI: case Op_CMoveL: case Op_CMoveF: case Op_CMoveD:
1458 *start = 2;
1459 *end = n->req();
1460 return;
1461 }
1462 *start = 1;
1463 *end = n->req(); // default is all operands
1464 }
1465
1466 //------------------------------in_packset---------------------------
1467 // Are s1 and s2 in a pack pair and ordered as s1,s2?
1468 bool SuperWord::in_packset(Node* s1, Node* s2) {
1469 for (int i = 0; i < _packset.length(); i++) {
1470 Node_List* p = _packset.at(i);
1471 assert(p->size() == 2, "must be");
1472 if (p->at(0) == s1 && p->at(p->size()-1) == s2) {
1473 return true;
1474 }
1475 }
1476 return false;
1477 }
1478
1479 //------------------------------in_pack---------------------------
1480 // Is s in pack p?
1481 Node_List* SuperWord::in_pack(Node* s, Node_List* p) {
1482 for (uint i = 0; i < p->size(); i++) {
1483 if (p->at(i) == s) {
1484 return p;
1485 }
1486 }
1487 return NULL;
1488 }
1489
1490 //------------------------------remove_pack_at---------------------------
1491 // Remove the pack at position pos in the packset
1492 void SuperWord::remove_pack_at(int pos) {
1493 Node_List* p = _packset.at(pos);
1494 for (uint i = 0; i < p->size(); i++) {
1495 Node* s = p->at(i);
1496 set_my_pack(s, NULL);
1497 }
1498 _packset.remove_at(pos);
1499 }
1500
1501 //------------------------------executed_first---------------------------
1502 // Return the node executed first in pack p. Uses the RPO block list
1503 // to determine order.
1504 Node* SuperWord::executed_first(Node_List* p) {
1505 Node* n = p->at(0);
1506 int n_rpo = bb_idx(n);
1507 for (uint i = 1; i < p->size(); i++) {
1508 Node* s = p->at(i);
1509 int s_rpo = bb_idx(s);
1510 if (s_rpo < n_rpo) {
1511 n = s;
1512 n_rpo = s_rpo;
1513 }
1514 }
1515 return n;
1516 }
1517
1518 //------------------------------executed_last---------------------------
1519 // Return the node executed last in pack p.
1520 Node* SuperWord::executed_last(Node_List* p) {
1521 Node* n = p->at(0);
1522 int n_rpo = bb_idx(n);
1523 for (uint i = 1; i < p->size(); i++) {
1524 Node* s = p->at(i);
1525 int s_rpo = bb_idx(s);
1526 if (s_rpo > n_rpo) {
1527 n = s;
1528 n_rpo = s_rpo;
1529 }
1530 }
1531 return n;
1532 }
1533
1534 //----------------------------align_initial_loop_index---------------------------
1535 // Adjust pre-loop limit so that in main loop, a load/store reference
1536 // to align_to_ref will be a position zero in the vector.
1537 // (iv + k) mod vector_align == 0
1538 void SuperWord::align_initial_loop_index(MemNode* align_to_ref) {
1539 CountedLoopNode *main_head = lp()->as_CountedLoop();
1540 assert(main_head->is_main_loop(), "");
1541 CountedLoopEndNode* pre_end = get_pre_loop_end(main_head);
1542 assert(pre_end != NULL, "");
1543 Node *pre_opaq1 = pre_end->limit();
1544 assert(pre_opaq1->Opcode() == Op_Opaque1, "");
1545 Opaque1Node *pre_opaq = (Opaque1Node*)pre_opaq1;
1546 Node *lim0 = pre_opaq->in(1);
1547
1548 // Where we put new limit calculations
1549 Node *pre_ctrl = pre_end->loopnode()->in(LoopNode::EntryControl);
1550
1551 // Ensure the original loop limit is available from the
1552 // pre-loop Opaque1 node.
1553 Node *orig_limit = pre_opaq->original_loop_limit();
1554 assert(orig_limit != NULL && _igvn.type(orig_limit) != Type::TOP, "");
1555
1556 SWPointer align_to_ref_p(align_to_ref, this);
1557
1558 // Given:
1559 // lim0 == original pre loop limit
1560 // V == v_align (power of 2)
1561 // invar == extra invariant piece of the address expression
1562 // e == k [ +/- invar ]
1563 //
1564 // When reassociating expressions involving '%' the basic rules are:
1565 // (a - b) % k == 0 => a % k == b % k
1566 // and:
1567 // (a + b) % k == 0 => a % k == (k - b) % k
1568 //
1569 // For stride > 0 && scale > 0,
1570 // Derive the new pre-loop limit "lim" such that the two constraints:
1571 // (1) lim = lim0 + N (where N is some positive integer < V)
1572 // (2) (e + lim) % V == 0
1573 // are true.
1574 //
1575 // Substituting (1) into (2),
1576 // (e + lim0 + N) % V == 0
1577 // solve for N:
1578 // N = (V - (e + lim0)) % V
1579 // substitute back into (1), so that new limit
1580 // lim = lim0 + (V - (e + lim0)) % V
1581 //
1582 // For stride > 0 && scale < 0
1583 // Constraints:
1584 // lim = lim0 + N
1585 // (e - lim) % V == 0
1586 // Solving for lim:
1587 // (e - lim0 - N) % V == 0
1588 // N = (e - lim0) % V
1589 // lim = lim0 + (e - lim0) % V
1590 //
1591 // For stride < 0 && scale > 0
1592 // Constraints:
1593 // lim = lim0 - N
1594 // (e + lim) % V == 0
1595 // Solving for lim:
1596 // (e + lim0 - N) % V == 0
1597 // N = (e + lim0) % V
1598 // lim = lim0 - (e + lim0) % V
1599 //
1600 // For stride < 0 && scale < 0
1601 // Constraints:
1602 // lim = lim0 - N
1603 // (e - lim) % V == 0
1604 // Solving for lim:
1605 // (e - lim0 + N) % V == 0
1606 // N = (V - (e - lim0)) % V
1607 // lim = lim0 - (V - (e - lim0)) % V
1608
1609 int stride = iv_stride();
1610 int scale = align_to_ref_p.scale_in_bytes();
1611 int elt_size = align_to_ref_p.memory_size();
1612 int v_align = vector_width_in_bytes() / elt_size;
1613 int k = align_to_ref_p.offset_in_bytes() / elt_size;
1614
1615 Node *kn = _igvn.intcon(k);
1616
1617 Node *e = kn;
1618 if (align_to_ref_p.invar() != NULL) {
1619 // incorporate any extra invariant piece producing k +/- invar >>> log2(elt)
1620 Node* log2_elt = _igvn.intcon(exact_log2(elt_size));
1621 Node* aref = new (_phase->C, 3) URShiftINode(align_to_ref_p.invar(), log2_elt);
1622 _phase->_igvn.register_new_node_with_optimizer(aref);
1623 _phase->set_ctrl(aref, pre_ctrl);
1624 if (align_to_ref_p.negate_invar()) {
1625 e = new (_phase->C, 3) SubINode(e, aref);
1626 } else {
1627 e = new (_phase->C, 3) AddINode(e, aref);
1628 }
1629 _phase->_igvn.register_new_node_with_optimizer(e);
1630 _phase->set_ctrl(e, pre_ctrl);
1631 }
1632
1633 // compute e +/- lim0
1634 if (scale < 0) {
1635 e = new (_phase->C, 3) SubINode(e, lim0);
1636 } else {
1637 e = new (_phase->C, 3) AddINode(e, lim0);
1638 }
1639 _phase->_igvn.register_new_node_with_optimizer(e);
1640 _phase->set_ctrl(e, pre_ctrl);
1641
1642 if (stride * scale > 0) {
1643 // compute V - (e +/- lim0)
1644 Node* va = _igvn.intcon(v_align);
1645 e = new (_phase->C, 3) SubINode(va, e);
1646 _phase->_igvn.register_new_node_with_optimizer(e);
1647 _phase->set_ctrl(e, pre_ctrl);
1648 }
1649 // compute N = (exp) % V
1650 Node* va_msk = _igvn.intcon(v_align - 1);
1651 Node* N = new (_phase->C, 3) AndINode(e, va_msk);
1652 _phase->_igvn.register_new_node_with_optimizer(N);
1653 _phase->set_ctrl(N, pre_ctrl);
1654
1655 // substitute back into (1), so that new limit
1656 // lim = lim0 + N
1657 Node* lim;
1658 if (stride < 0) {
1659 lim = new (_phase->C, 3) SubINode(lim0, N);
1660 } else {
1661 lim = new (_phase->C, 3) AddINode(lim0, N);
1662 }
1663 _phase->_igvn.register_new_node_with_optimizer(lim);
1664 _phase->set_ctrl(lim, pre_ctrl);
1665 Node* constrained =
1666 (stride > 0) ? (Node*) new (_phase->C,3) MinINode(lim, orig_limit)
1667 : (Node*) new (_phase->C,3) MaxINode(lim, orig_limit);
1668 _phase->_igvn.register_new_node_with_optimizer(constrained);
1669 _phase->set_ctrl(constrained, pre_ctrl);
1670 _igvn.hash_delete(pre_opaq);
1671 pre_opaq->set_req(1, constrained);
1672 }
1673
1674 //----------------------------get_pre_loop_end---------------------------
1675 // Find pre loop end from main loop. Returns null if none.
1676 CountedLoopEndNode* SuperWord::get_pre_loop_end(CountedLoopNode *cl) {
1677 Node *ctrl = cl->in(LoopNode::EntryControl);
1678 if (!ctrl->is_IfTrue() && !ctrl->is_IfFalse()) return NULL;
1679 Node *iffm = ctrl->in(0);
1680 if (!iffm->is_If()) return NULL;
1681 Node *p_f = iffm->in(0);
1682 if (!p_f->is_IfFalse()) return NULL;
1683 if (!p_f->in(0)->is_CountedLoopEnd()) return NULL;
1684 CountedLoopEndNode *pre_end = p_f->in(0)->as_CountedLoopEnd();
1685 if (!pre_end->loopnode()->is_pre_loop()) return NULL;
1686 return pre_end;
1687 }
1688
1689
1690 //------------------------------init---------------------------
1691 void SuperWord::init() {
1692 _dg.init();
1693 _packset.clear();
1694 _disjoint_ptrs.clear();
1695 _block.clear();
1696 _data_entry.clear();
1697 _mem_slice_head.clear();
1698 _mem_slice_tail.clear();
1699 _node_info.clear();
1700 _align_to_ref = NULL;
1701 _lpt = NULL;
1702 _lp = NULL;
1703 _bb = NULL;
1704 _iv = NULL;
1705 }
1706
1707 //------------------------------print_packset---------------------------
1708 void SuperWord::print_packset() {
1709 #ifndef PRODUCT
1710 tty->print_cr("packset");
1711 for (int i = 0; i < _packset.length(); i++) {
1712 tty->print_cr("Pack: %d", i);
1713 Node_List* p = _packset.at(i);
1714 print_pack(p);
1715 }
1716 #endif
1717 }
1718
1719 //------------------------------print_pack---------------------------
1720 void SuperWord::print_pack(Node_List* p) {
1721 for (uint i = 0; i < p->size(); i++) {
1722 print_stmt(p->at(i));
1723 }
1724 }
1725
1726 //------------------------------print_bb---------------------------
1727 void SuperWord::print_bb() {
1728 #ifndef PRODUCT
1729 tty->print_cr("\nBlock");
1730 for (int i = 0; i < _block.length(); i++) {
1731 Node* n = _block.at(i);
1732 tty->print("%d ", i);
1733 if (n) {
1734 n->dump();
1735 }
1736 }
1737 #endif
1738 }
1739
1740 //------------------------------print_stmt---------------------------
1741 void SuperWord::print_stmt(Node* s) {
1742 #ifndef PRODUCT
1743 tty->print(" align: %d \t", alignment(s));
1744 s->dump();
1745 #endif
1746 }
1747
1748 //------------------------------blank---------------------------
1749 char* SuperWord::blank(uint depth) {
1750 static char blanks[101];
1751 assert(depth < 101, "too deep");
1752 for (uint i = 0; i < depth; i++) blanks[i] = ' ';
1753 blanks[depth] = '\0';
1754 return blanks;
1755 }
1756
1757
1758 //==============================SWPointer===========================
1759
1760 //----------------------------SWPointer------------------------
1761 SWPointer::SWPointer(MemNode* mem, SuperWord* slp) :
1762 _mem(mem), _slp(slp), _base(NULL), _adr(NULL),
1763 _scale(0), _offset(0), _invar(NULL), _negate_invar(false) {
1764
1765 Node* adr = mem->in(MemNode::Address);
1766 if (!adr->is_AddP()) {
1767 assert(!valid(), "too complex");
1768 return;
1769 }
1770 // Match AddP(base, AddP(ptr, k*iv [+ invariant]), constant)
1771 Node* base = adr->in(AddPNode::Base);
1772 for (int i = 0; i < 3; i++) {
1773 if (!scaled_iv_plus_offset(adr->in(AddPNode::Offset))) {
1774 assert(!valid(), "too complex");
1775 return;
1776 }
1777 adr = adr->in(AddPNode::Address);
1778 if (base == adr || !adr->is_AddP()) {
1779 break; // stop looking at addp's
1780 }
1781 }
1782 _base = base;
1783 _adr = adr;
1784 assert(valid(), "Usable");
1785 }
1786
1787 // Following is used to create a temporary object during
1788 // the pattern match of an address expression.
1789 SWPointer::SWPointer(SWPointer* p) :
1790 _mem(p->_mem), _slp(p->_slp), _base(NULL), _adr(NULL),
1791 _scale(0), _offset(0), _invar(NULL), _negate_invar(false) {}
1792
1793 //------------------------scaled_iv_plus_offset--------------------
1794 // Match: k*iv + offset
1795 // where: k is a constant that maybe zero, and
1796 // offset is (k2 [+/- invariant]) where k2 maybe zero and invariant is optional
1797 bool SWPointer::scaled_iv_plus_offset(Node* n) {
1798 if (scaled_iv(n)) {
1799 return true;
1800 }
1801 if (offset_plus_k(n)) {
1802 return true;
1803 }
1804 int opc = n->Opcode();
1805 if (opc == Op_AddI) {
1806 if (scaled_iv(n->in(1)) && offset_plus_k(n->in(2))) {
1807 return true;
1808 }
1809 if (scaled_iv(n->in(2)) && offset_plus_k(n->in(1))) {
1810 return true;
1811 }
1812 } else if (opc == Op_SubI) {
1813 if (scaled_iv(n->in(1)) && offset_plus_k(n->in(2), true)) {
1814 return true;
1815 }
1816 if (scaled_iv(n->in(2)) && offset_plus_k(n->in(1))) {
1817 _scale *= -1;
1818 return true;
1819 }
1820 }
1821 return false;
1822 }
1823
1824 //----------------------------scaled_iv------------------------
1825 // Match: k*iv where k is a constant that's not zero
1826 bool SWPointer::scaled_iv(Node* n) {
1827 if (_scale != 0) {
1828 return false; // already found a scale
1829 }
1830 if (n == iv()) {
1831 _scale = 1;
1832 return true;
1833 }
1834 int opc = n->Opcode();
1835 if (opc == Op_MulI) {
1836 if (n->in(1) == iv() && n->in(2)->is_Con()) {
1837 _scale = n->in(2)->get_int();
1838 return true;
1839 } else if (n->in(2) == iv() && n->in(1)->is_Con()) {
1840 _scale = n->in(1)->get_int();
1841 return true;
1842 }
1843 } else if (opc == Op_LShiftI) {
1844 if (n->in(1) == iv() && n->in(2)->is_Con()) {
1845 _scale = 1 << n->in(2)->get_int();
1846 return true;
1847 }
1848 } else if (opc == Op_ConvI2L) {
1849 if (scaled_iv_plus_offset(n->in(1))) {
1850 return true;
1851 }
1852 } else if (opc == Op_LShiftL) {
1853 if (!has_iv() && _invar == NULL) {
1854 // Need to preserve the current _offset value, so
1855 // create a temporary object for this expression subtree.
1856 // Hacky, so should re-engineer the address pattern match.
1857 SWPointer tmp(this);
1858 if (tmp.scaled_iv_plus_offset(n->in(1))) {
1859 if (tmp._invar == NULL) {
1860 int mult = 1 << n->in(2)->get_int();
1861 _scale = tmp._scale * mult;
1862 _offset += tmp._offset * mult;
1863 return true;
1864 }
1865 }
1866 }
1867 }
1868 return false;
1869 }
1870
1871 //----------------------------offset_plus_k------------------------
1872 // Match: offset is (k [+/- invariant])
1873 // where k maybe zero and invariant is optional, but not both.
1874 bool SWPointer::offset_plus_k(Node* n, bool negate) {
1875 int opc = n->Opcode();
1876 if (opc == Op_ConI) {
1877 _offset += negate ? -(n->get_int()) : n->get_int();
1878 return true;
1879 } else if (opc == Op_ConL) {
1880 // Okay if value fits into an int
1881 const TypeLong* t = n->find_long_type();
1882 if (t->higher_equal(TypeLong::INT)) {
1883 jlong loff = n->get_long();
1884 jint off = (jint)loff;
1885 _offset += negate ? -off : loff;
1886 return true;
1887 }
1888 return false;
1889 }
1890 if (_invar != NULL) return false; // already have an invariant
1891 if (opc == Op_AddI) {
1892 if (n->in(2)->is_Con() && invariant(n->in(1))) {
1893 _negate_invar = negate;
1894 _invar = n->in(1);
1895 _offset += negate ? -(n->in(2)->get_int()) : n->in(2)->get_int();
1896 return true;
1897 } else if (n->in(1)->is_Con() && invariant(n->in(2))) {
1898 _offset += negate ? -(n->in(1)->get_int()) : n->in(1)->get_int();
1899 _negate_invar = negate;
1900 _invar = n->in(2);
1901 return true;
1902 }
1903 }
1904 if (opc == Op_SubI) {
1905 if (n->in(2)->is_Con() && invariant(n->in(1))) {
1906 _negate_invar = negate;
1907 _invar = n->in(1);
1908 _offset += !negate ? -(n->in(2)->get_int()) : n->in(2)->get_int();
1909 return true;
1910 } else if (n->in(1)->is_Con() && invariant(n->in(2))) {
1911 _offset += negate ? -(n->in(1)->get_int()) : n->in(1)->get_int();
1912 _negate_invar = !negate;
1913 _invar = n->in(2);
1914 return true;
1915 }
1916 }
1917 if (invariant(n)) {
1918 _negate_invar = negate;
1919 _invar = n;
1920 return true;
1921 }
1922 return false;
1923 }
1924
1925 //----------------------------print------------------------
1926 void SWPointer::print() {
1927 #ifndef PRODUCT
1928 tty->print("base: %d adr: %d scale: %d offset: %d invar: %c%d\n",
1929 _base != NULL ? _base->_idx : 0,
1930 _adr != NULL ? _adr->_idx : 0,
1931 _scale, _offset,
1932 _negate_invar?'-':'+',
1933 _invar != NULL ? _invar->_idx : 0);
1934 #endif
1935 }
1936
1937 // ========================= OrderedPair =====================
1938
1939 const OrderedPair OrderedPair::initial;
1940
1941 // ========================= SWNodeInfo =====================
1942
1943 const SWNodeInfo SWNodeInfo::initial;
1944
1945
1946 // ============================ DepGraph ===========================
1947
1948 //------------------------------make_node---------------------------
1949 // Make a new dependence graph node for an ideal node.
1950 DepMem* DepGraph::make_node(Node* node) {
1951 DepMem* m = new (_arena) DepMem(node);
1952 if (node != NULL) {
1953 assert(_map.at_grow(node->_idx) == NULL, "one init only");
1954 _map.at_put_grow(node->_idx, m);
1955 }
1956 return m;
1957 }
1958
1959 //------------------------------make_edge---------------------------
1960 // Make a new dependence graph edge from dpred -> dsucc
1961 DepEdge* DepGraph::make_edge(DepMem* dpred, DepMem* dsucc) {
1962 DepEdge* e = new (_arena) DepEdge(dpred, dsucc, dsucc->in_head(), dpred->out_head());
1963 dpred->set_out_head(e);
1964 dsucc->set_in_head(e);
1965 return e;
1966 }
1967
1968 // ========================== DepMem ========================
1969
1970 //------------------------------in_cnt---------------------------
1971 int DepMem::in_cnt() {
1972 int ct = 0;
1973 for (DepEdge* e = _in_head; e != NULL; e = e->next_in()) ct++;
1974 return ct;
1975 }
1976
1977 //------------------------------out_cnt---------------------------
1978 int DepMem::out_cnt() {
1979 int ct = 0;
1980 for (DepEdge* e = _out_head; e != NULL; e = e->next_out()) ct++;
1981 return ct;
1982 }
1983
1984 //------------------------------print-----------------------------
1985 void DepMem::print() {
1986 #ifndef PRODUCT
1987 tty->print(" DepNode %d (", _node->_idx);
1988 for (DepEdge* p = _in_head; p != NULL; p = p->next_in()) {
1989 Node* pred = p->pred()->node();
1990 tty->print(" %d", pred != NULL ? pred->_idx : 0);
1991 }
1992 tty->print(") [");
1993 for (DepEdge* s = _out_head; s != NULL; s = s->next_out()) {
1994 Node* succ = s->succ()->node();
1995 tty->print(" %d", succ != NULL ? succ->_idx : 0);
1996 }
1997 tty->print_cr(" ]");
1998 #endif
1999 }
2000
2001 // =========================== DepEdge =========================
2002
2003 //------------------------------DepPreds---------------------------
2004 void DepEdge::print() {
2005 #ifndef PRODUCT
2006 tty->print_cr("DepEdge: %d [ %d ]", _pred->node()->_idx, _succ->node()->_idx);
2007 #endif
2008 }
2009
2010 // =========================== DepPreds =========================
2011 // Iterator over predecessor edges in the dependence graph.
2012
2013 //------------------------------DepPreds---------------------------
2014 DepPreds::DepPreds(Node* n, DepGraph& dg) {
2015 _n = n;
2016 _done = false;
2017 if (_n->is_Store() || _n->is_Load()) {
2018 _next_idx = MemNode::Address;
2019 _end_idx = n->req();
2020 _dep_next = dg.dep(_n)->in_head();
2021 } else if (_n->is_Mem()) {
2022 _next_idx = 0;
2023 _end_idx = 0;
2024 _dep_next = dg.dep(_n)->in_head();
2025 } else {
2026 _next_idx = 1;
2027 _end_idx = _n->req();
2028 _dep_next = NULL;
2029 }
2030 next();
2031 }
2032
2033 //------------------------------next---------------------------
2034 void DepPreds::next() {
2035 if (_dep_next != NULL) {
2036 _current = _dep_next->pred()->node();
2037 _dep_next = _dep_next->next_in();
2038 } else if (_next_idx < _end_idx) {
2039 _current = _n->in(_next_idx++);
2040 } else {
2041 _done = true;
2042 }
2043 }
2044
2045 // =========================== DepSuccs =========================
2046 // Iterator over successor edges in the dependence graph.
2047
2048 //------------------------------DepSuccs---------------------------
2049 DepSuccs::DepSuccs(Node* n, DepGraph& dg) {
2050 _n = n;
2051 _done = false;
2052 if (_n->is_Load()) {
2053 _next_idx = 0;
2054 _end_idx = _n->outcnt();
2055 _dep_next = dg.dep(_n)->out_head();
2056 } else if (_n->is_Mem() || _n->is_Phi() && _n->bottom_type() == Type::MEMORY) {
2057 _next_idx = 0;
2058 _end_idx = 0;
2059 _dep_next = dg.dep(_n)->out_head();
2060 } else {
2061 _next_idx = 0;
2062 _end_idx = _n->outcnt();
2063 _dep_next = NULL;
2064 }
2065 next();
2066 }
2067
2068 //-------------------------------next---------------------------
2069 void DepSuccs::next() {
2070 if (_dep_next != NULL) {
2071 _current = _dep_next->succ()->node();
2072 _dep_next = _dep_next->next_out();
2073 } else if (_next_idx < _end_idx) {
2074 _current = _n->raw_out(_next_idx++);
2075 } else {
2076 _done = true;
2077 }
2078 }