1 /*
2 * Copyright 2007-2009 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 // Make sure the are no extra control users of the loop backedge
69 if (cl->back_control()->outcnt() != 1) {
70 return;
71 }
72
73 // Check for pre-loop ending with CountedLoopEnd(Bool(Cmp(x,Opaque1(limit))))
74 CountedLoopEndNode* pre_end = get_pre_loop_end(cl);
75 if (pre_end == NULL) return;
76 Node *pre_opaq1 = pre_end->limit();
77 if (pre_opaq1->Opcode() != Op_Opaque1) return;
78
79 // Do vectors exist on this architecture?
80 if (vector_width_in_bytes() == 0) return;
81
82 init(); // initialize data structures
83
84 set_lpt(lpt);
85 set_lp(cl);
86
87 // For now, define one block which is the entire loop body
88 set_bb(cl);
89
90 assert(_packset.length() == 0, "packset must be empty");
91 SLP_extract();
92 }
93
94 //------------------------------SLP_extract---------------------------
95 // Extract the superword level parallelism
96 //
97 // 1) A reverse post-order of nodes in the block is constructed. By scanning
98 // this list from first to last, all definitions are visited before their uses.
99 //
100 // 2) A point-to-point dependence graph is constructed between memory references.
101 // This simplies the upcoming "independence" checker.
102 //
103 // 3) The maximum depth in the node graph from the beginning of the block
104 // to each node is computed. This is used to prune the graph search
105 // in the independence checker.
106 //
107 // 4) For integer types, the necessary bit width is propagated backwards
108 // from stores to allow packed operations on byte, char, and short
109 // integers. This reverses the promotion to type "int" that javac
110 // did for operations like: char c1,c2,c3; c1 = c2 + c3.
111 //
112 // 5) One of the memory references is picked to be an aligned vector reference.
113 // The pre-loop trip count is adjusted to align this reference in the
114 // unrolled body.
115 //
116 // 6) The initial set of pack pairs is seeded with memory references.
117 //
118 // 7) The set of pack pairs is extended by following use->def and def->use links.
119 //
120 // 8) The pairs are combined into vector sized packs.
121 //
122 // 9) Reorder the memory slices to co-locate members of the memory packs.
123 //
124 // 10) Generate ideal vector nodes for the final set of packs and where necessary,
125 // inserting scalar promotion, vector creation from multiple scalars, and
126 // extraction of scalar values from vectors.
127 //
128 void SuperWord::SLP_extract() {
129
130 // Ready the block
131
132 construct_bb();
133
134 dependence_graph();
135
136 compute_max_depth();
137
138 compute_vector_element_type();
139
140 // Attempt vectorization
141
142 find_adjacent_refs();
143
144 extend_packlist();
145
146 combine_packs();
147
148 construct_my_pack_map();
149
150 filter_packs();
151
152 schedule();
153
154 output();
155 }
156
157 //------------------------------find_adjacent_refs---------------------------
158 // Find the adjacent memory references and create pack pairs for them.
159 // This is the initial set of packs that will then be extended by
160 // following use->def and def->use links. The align positions are
161 // assigned relative to the reference "align_to_ref"
162 void SuperWord::find_adjacent_refs() {
163 // Get list of memory operations
164 Node_List memops;
165 for (int i = 0; i < _block.length(); i++) {
166 Node* n = _block.at(i);
167 if (n->is_Mem() && in_bb(n) &&
168 is_java_primitive(n->as_Mem()->memory_type())) {
169 int align = memory_alignment(n->as_Mem(), 0);
170 if (align != bottom_align) {
171 memops.push(n);
172 }
173 }
174 }
175 if (memops.size() == 0) return;
176
177 // Find a memory reference to align to. The pre-loop trip count
178 // is modified to align this reference to a vector-aligned address
179 find_align_to_ref(memops);
180 if (align_to_ref() == NULL) return;
181
182 SWPointer align_to_ref_p(align_to_ref(), this);
183 int offset = align_to_ref_p.offset_in_bytes();
184 int scale = align_to_ref_p.scale_in_bytes();
185 int vw = vector_width_in_bytes();
186 int stride_sign = (scale * iv_stride()) > 0 ? 1 : -1;
187 int iv_adjustment = (stride_sign * vw - (offset % vw)) % vw;
188
189 #ifndef PRODUCT
190 if (TraceSuperWord)
191 tty->print_cr("\noffset = %d iv_adjustment = %d elt_align = %d scale = %d iv_stride = %d",
192 offset, iv_adjustment, align_to_ref_p.memory_size(), align_to_ref_p.scale_in_bytes(), iv_stride());
193 #endif
194
195 // Set alignment relative to "align_to_ref"
196 for (int i = memops.size() - 1; i >= 0; i--) {
197 MemNode* s = memops.at(i)->as_Mem();
198 SWPointer p2(s, this);
199 if (p2.comparable(align_to_ref_p)) {
200 int align = memory_alignment(s, iv_adjustment);
201 set_alignment(s, align);
202 } else {
203 memops.remove(i);
204 }
205 }
206
207 // Create initial pack pairs of memory operations
208 for (uint i = 0; i < memops.size(); i++) {
209 Node* s1 = memops.at(i);
210 for (uint j = 0; j < memops.size(); j++) {
211 Node* s2 = memops.at(j);
212 if (s1 != s2 && are_adjacent_refs(s1, s2)) {
213 int align = alignment(s1);
214 if (stmts_can_pack(s1, s2, align)) {
215 Node_List* pair = new Node_List();
216 pair->push(s1);
217 pair->push(s2);
218 _packset.append(pair);
219 }
220 }
221 }
222 }
223
224 #ifndef PRODUCT
225 if (TraceSuperWord) {
226 tty->print_cr("\nAfter find_adjacent_refs");
227 print_packset();
228 }
229 #endif
230 }
231
232 //------------------------------find_align_to_ref---------------------------
233 // Find a memory reference to align the loop induction variable to.
234 // Looks first at stores then at loads, looking for a memory reference
235 // with the largest number of references similar to it.
236 void SuperWord::find_align_to_ref(Node_List &memops) {
237 GrowableArray<int> cmp_ct(arena(), memops.size(), memops.size(), 0);
238
239 // Count number of comparable memory ops
240 for (uint i = 0; i < memops.size(); i++) {
241 MemNode* s1 = memops.at(i)->as_Mem();
242 SWPointer p1(s1, this);
243 // Discard if pre loop can't align this reference
244 if (!ref_is_alignable(p1)) {
245 *cmp_ct.adr_at(i) = 0;
246 continue;
247 }
248 for (uint j = i+1; j < memops.size(); j++) {
249 MemNode* s2 = memops.at(j)->as_Mem();
250 if (isomorphic(s1, s2)) {
251 SWPointer p2(s2, this);
252 if (p1.comparable(p2)) {
253 (*cmp_ct.adr_at(i))++;
254 (*cmp_ct.adr_at(j))++;
255 }
256 }
257 }
258 }
259
260 // Find Store (or Load) with the greatest number of "comparable" references
261 int max_ct = 0;
262 int max_idx = -1;
263 int min_size = max_jint;
264 int min_iv_offset = max_jint;
265 for (uint j = 0; j < memops.size(); j++) {
266 MemNode* s = memops.at(j)->as_Mem();
267 if (s->is_Store()) {
268 SWPointer p(s, this);
269 if (cmp_ct.at(j) > max_ct ||
270 cmp_ct.at(j) == max_ct && (data_size(s) < min_size ||
271 data_size(s) == min_size &&
272 p.offset_in_bytes() < min_iv_offset)) {
273 max_ct = cmp_ct.at(j);
274 max_idx = j;
275 min_size = data_size(s);
276 min_iv_offset = p.offset_in_bytes();
277 }
278 }
279 }
280 // If no stores, look at loads
281 if (max_ct == 0) {
282 for (uint j = 0; j < memops.size(); j++) {
283 MemNode* s = memops.at(j)->as_Mem();
284 if (s->is_Load()) {
285 SWPointer p(s, this);
286 if (cmp_ct.at(j) > max_ct ||
287 cmp_ct.at(j) == max_ct && (data_size(s) < min_size ||
288 data_size(s) == min_size &&
289 p.offset_in_bytes() < min_iv_offset)) {
290 max_ct = cmp_ct.at(j);
291 max_idx = j;
292 min_size = data_size(s);
293 min_iv_offset = p.offset_in_bytes();
294 }
295 }
296 }
297 }
298
299 if (max_ct > 0)
300 set_align_to_ref(memops.at(max_idx)->as_Mem());
301
302 #ifndef PRODUCT
303 if (TraceSuperWord && Verbose) {
304 tty->print_cr("\nVector memops after find_align_to_refs");
305 for (uint i = 0; i < memops.size(); i++) {
306 MemNode* s = memops.at(i)->as_Mem();
307 s->dump();
308 }
309 }
310 #endif
311 }
312
313 //------------------------------ref_is_alignable---------------------------
314 // Can the preloop align the reference to position zero in the vector?
315 bool SuperWord::ref_is_alignable(SWPointer& p) {
316 if (!p.has_iv()) {
317 return true; // no induction variable
318 }
319 CountedLoopEndNode* pre_end = get_pre_loop_end(lp()->as_CountedLoop());
320 assert(pre_end->stride_is_con(), "pre loop stride is constant");
321 int preloop_stride = pre_end->stride_con();
322
323 int span = preloop_stride * p.scale_in_bytes();
324
325 // Stride one accesses are alignable.
326 if (ABS(span) == p.memory_size())
327 return true;
328
329 // If initial offset from start of object is computable,
330 // compute alignment within the vector.
331 int vw = vector_width_in_bytes();
332 if (vw % span == 0) {
333 Node* init_nd = pre_end->init_trip();
334 if (init_nd->is_Con() && p.invar() == NULL) {
335 int init = init_nd->bottom_type()->is_int()->get_con();
336
337 int init_offset = init * p.scale_in_bytes() + p.offset_in_bytes();
338 assert(init_offset >= 0, "positive offset from object start");
339
340 if (span > 0) {
341 return (vw - (init_offset % vw)) % span == 0;
342 } else {
343 assert(span < 0, "nonzero stride * scale");
344 return (init_offset % vw) % -span == 0;
345 }
346 }
347 }
348 return false;
349 }
350
351 //---------------------------dependence_graph---------------------------
352 // Construct dependency graph.
353 // Add dependence edges to load/store nodes for memory dependence
354 // A.out()->DependNode.in(1) and DependNode.out()->B.prec(x)
355 void SuperWord::dependence_graph() {
356 // First, assign a dependence node to each memory node
357 for (int i = 0; i < _block.length(); i++ ) {
358 Node *n = _block.at(i);
359 if (n->is_Mem() || n->is_Phi() && n->bottom_type() == Type::MEMORY) {
360 _dg.make_node(n);
361 }
362 }
363
364 // For each memory slice, create the dependences
365 for (int i = 0; i < _mem_slice_head.length(); i++) {
366 Node* n = _mem_slice_head.at(i);
367 Node* n_tail = _mem_slice_tail.at(i);
368
369 // Get slice in predecessor order (last is first)
370 mem_slice_preds(n_tail, n, _nlist);
371
372 // Make the slice dependent on the root
373 DepMem* slice = _dg.dep(n);
374 _dg.make_edge(_dg.root(), slice);
375
376 // Create a sink for the slice
377 DepMem* slice_sink = _dg.make_node(NULL);
378 _dg.make_edge(slice_sink, _dg.tail());
379
380 // Now visit each pair of memory ops, creating the edges
381 for (int j = _nlist.length() - 1; j >= 0 ; j--) {
382 Node* s1 = _nlist.at(j);
383
384 // If no dependency yet, use slice
385 if (_dg.dep(s1)->in_cnt() == 0) {
386 _dg.make_edge(slice, s1);
387 }
388 SWPointer p1(s1->as_Mem(), this);
389 bool sink_dependent = true;
390 for (int k = j - 1; k >= 0; k--) {
391 Node* s2 = _nlist.at(k);
392 if (s1->is_Load() && s2->is_Load())
393 continue;
394 SWPointer p2(s2->as_Mem(), this);
395
396 int cmp = p1.cmp(p2);
397 if (SuperWordRTDepCheck &&
398 p1.base() != p2.base() && p1.valid() && p2.valid()) {
399 // Create a runtime check to disambiguate
400 OrderedPair pp(p1.base(), p2.base());
401 _disjoint_ptrs.append_if_missing(pp);
402 } else if (!SWPointer::not_equal(cmp)) {
403 // Possibly same address
404 _dg.make_edge(s1, s2);
405 sink_dependent = false;
406 }
407 }
408 if (sink_dependent) {
409 _dg.make_edge(s1, slice_sink);
410 }
411 }
412 #ifndef PRODUCT
413 if (TraceSuperWord) {
414 tty->print_cr("\nDependence graph for slice: %d", n->_idx);
415 for (int q = 0; q < _nlist.length(); q++) {
416 _dg.print(_nlist.at(q));
417 }
418 tty->cr();
419 }
420 #endif
421 _nlist.clear();
422 }
423
424 #ifndef PRODUCT
425 if (TraceSuperWord) {
426 tty->print_cr("\ndisjoint_ptrs: %s", _disjoint_ptrs.length() > 0 ? "" : "NONE");
427 for (int r = 0; r < _disjoint_ptrs.length(); r++) {
428 _disjoint_ptrs.at(r).print();
429 tty->cr();
430 }
431 tty->cr();
432 }
433 #endif
434 }
435
436 //---------------------------mem_slice_preds---------------------------
437 // Return a memory slice (node list) in predecessor order starting at "start"
438 void SuperWord::mem_slice_preds(Node* start, Node* stop, GrowableArray<Node*> &preds) {
439 assert(preds.length() == 0, "start empty");
440 Node* n = start;
441 Node* prev = NULL;
442 while (true) {
443 assert(in_bb(n), "must be in block");
444 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
445 Node* out = n->fast_out(i);
446 if (out->is_Load()) {
447 if (in_bb(out)) {
448 preds.push(out);
449 }
450 } else {
451 // FIXME
452 if (out->is_MergeMem() && !in_bb(out)) {
453 // Either unrolling is causing a memory edge not to disappear,
454 // or need to run igvn.optimize() again before SLP
455 } else if (out->is_Phi() && out->bottom_type() == Type::MEMORY && !in_bb(out)) {
456 // Ditto. Not sure what else to check further.
457 } else if (out->Opcode() == Op_StoreCM && out->in(MemNode::OopStore) == n) {
458 // StoreCM has an input edge used as a precedence edge.
459 // Maybe an issue when oop stores are vectorized.
460 } else if( out->is_MergeMem() && prev &&
461 prev->Opcode() == Op_StoreCM && out == prev->in(MemNode::OopStore)) {
462 // Oop store is a MergeMem! This should not happen. Temporarily remove the assertion
463 // for this case because it could not be superwordized anyway.
464 } else {
465 assert(out == prev || prev == NULL, "no branches off of store slice");
466 }
467 }
468 }
469 if (n == stop) break;
470 preds.push(n);
471 prev = n;
472 n = n->in(MemNode::Memory);
473 }
474 }
475
476 //------------------------------stmts_can_pack---------------------------
477 // Can s1 and s2 be in a pack with s1 immediately preceding s2 and
478 // s1 aligned at "align"
479 bool SuperWord::stmts_can_pack(Node* s1, Node* s2, int align) {
480 if (isomorphic(s1, s2)) {
481 if (independent(s1, s2)) {
482 if (!exists_at(s1, 0) && !exists_at(s2, 1)) {
483 if (!s1->is_Mem() || are_adjacent_refs(s1, s2)) {
484 int s1_align = alignment(s1);
485 int s2_align = alignment(s2);
486 if (s1_align == top_align || s1_align == align) {
487 if (s2_align == top_align || s2_align == align + data_size(s1)) {
488 return true;
489 }
490 }
491 }
492 }
493 }
494 }
495 return false;
496 }
497
498 //------------------------------exists_at---------------------------
499 // Does s exist in a pack at position pos?
500 bool SuperWord::exists_at(Node* s, uint pos) {
501 for (int i = 0; i < _packset.length(); i++) {
502 Node_List* p = _packset.at(i);
503 if (p->at(pos) == s) {
504 return true;
505 }
506 }
507 return false;
508 }
509
510 //------------------------------are_adjacent_refs---------------------------
511 // Is s1 immediately before s2 in memory?
512 bool SuperWord::are_adjacent_refs(Node* s1, Node* s2) {
513 if (!s1->is_Mem() || !s2->is_Mem()) return false;
514 if (!in_bb(s1) || !in_bb(s2)) return false;
515 // FIXME - co_locate_pack fails on Stores in different mem-slices, so
516 // only pack memops that are in the same alias set until that's fixed.
517 if (_phase->C->get_alias_index(s1->as_Mem()->adr_type()) !=
518 _phase->C->get_alias_index(s2->as_Mem()->adr_type()))
519 return false;
520 SWPointer p1(s1->as_Mem(), this);
521 SWPointer p2(s2->as_Mem(), this);
522 if (p1.base() != p2.base() || !p1.comparable(p2)) return false;
523 int diff = p2.offset_in_bytes() - p1.offset_in_bytes();
524 return diff == data_size(s1);
525 }
526
527 //------------------------------isomorphic---------------------------
528 // Are s1 and s2 similar?
529 bool SuperWord::isomorphic(Node* s1, Node* s2) {
530 if (s1->Opcode() != s2->Opcode()) return false;
531 if (s1->req() != s2->req()) return false;
532 if (s1->in(0) != s2->in(0)) return false;
533 if (velt_type(s1) != velt_type(s2)) return false;
534 return true;
535 }
536
537 //------------------------------independent---------------------------
538 // Is there no data path from s1 to s2 or s2 to s1?
539 bool SuperWord::independent(Node* s1, Node* s2) {
540 // assert(s1->Opcode() == s2->Opcode(), "check isomorphic first");
541 int d1 = depth(s1);
542 int d2 = depth(s2);
543 if (d1 == d2) return s1 != s2;
544 Node* deep = d1 > d2 ? s1 : s2;
545 Node* shallow = d1 > d2 ? s2 : s1;
546
547 visited_clear();
548
549 return independent_path(shallow, deep);
550 }
551
552 //------------------------------independent_path------------------------------
553 // Helper for independent
554 bool SuperWord::independent_path(Node* shallow, Node* deep, uint dp) {
555 if (dp >= 1000) return false; // stop deep recursion
556 visited_set(deep);
557 int shal_depth = depth(shallow);
558 assert(shal_depth <= depth(deep), "must be");
559 for (DepPreds preds(deep, _dg); !preds.done(); preds.next()) {
560 Node* pred = preds.current();
561 if (in_bb(pred) && !visited_test(pred)) {
562 if (shallow == pred) {
563 return false;
564 }
565 if (shal_depth < depth(pred) && !independent_path(shallow, pred, dp+1)) {
566 return false;
567 }
568 }
569 }
570 return true;
571 }
572
573 //------------------------------set_alignment---------------------------
574 void SuperWord::set_alignment(Node* s1, Node* s2, int align) {
575 set_alignment(s1, align);
576 set_alignment(s2, align + data_size(s1));
577 }
578
579 //------------------------------data_size---------------------------
580 int SuperWord::data_size(Node* s) {
581 const Type* t = velt_type(s);
582 BasicType bt = t->array_element_basic_type();
583 int bsize = type2aelembytes(bt);
584 assert(bsize != 0, "valid size");
585 return bsize;
586 }
587
588 //------------------------------extend_packlist---------------------------
589 // Extend packset by following use->def and def->use links from pack members.
590 void SuperWord::extend_packlist() {
591 bool changed;
592 do {
593 changed = false;
594 for (int i = 0; i < _packset.length(); i++) {
595 Node_List* p = _packset.at(i);
596 changed |= follow_use_defs(p);
597 changed |= follow_def_uses(p);
598 }
599 } while (changed);
600
601 #ifndef PRODUCT
602 if (TraceSuperWord) {
603 tty->print_cr("\nAfter extend_packlist");
604 print_packset();
605 }
606 #endif
607 }
608
609 //------------------------------follow_use_defs---------------------------
610 // Extend the packset by visiting operand definitions of nodes in pack p
611 bool SuperWord::follow_use_defs(Node_List* p) {
612 Node* s1 = p->at(0);
613 Node* s2 = p->at(1);
614 assert(p->size() == 2, "just checking");
615 assert(s1->req() == s2->req(), "just checking");
616 assert(alignment(s1) + data_size(s1) == alignment(s2), "just checking");
617
618 if (s1->is_Load()) return false;
619
620 int align = alignment(s1);
621 bool changed = false;
622 int start = s1->is_Store() ? MemNode::ValueIn : 1;
623 int end = s1->is_Store() ? MemNode::ValueIn+1 : s1->req();
624 for (int j = start; j < end; j++) {
625 Node* t1 = s1->in(j);
626 Node* t2 = s2->in(j);
627 if (!in_bb(t1) || !in_bb(t2))
628 continue;
629 if (stmts_can_pack(t1, t2, align)) {
630 if (est_savings(t1, t2) >= 0) {
631 Node_List* pair = new Node_List();
632 pair->push(t1);
633 pair->push(t2);
634 _packset.append(pair);
635 set_alignment(t1, t2, align);
636 changed = true;
637 }
638 }
639 }
640 return changed;
641 }
642
643 //------------------------------follow_def_uses---------------------------
644 // Extend the packset by visiting uses of nodes in pack p
645 bool SuperWord::follow_def_uses(Node_List* p) {
646 bool changed = false;
647 Node* s1 = p->at(0);
648 Node* s2 = p->at(1);
649 assert(p->size() == 2, "just checking");
650 assert(s1->req() == s2->req(), "just checking");
651 assert(alignment(s1) + data_size(s1) == alignment(s2), "just checking");
652
653 if (s1->is_Store()) return false;
654
655 int align = alignment(s1);
656 int savings = -1;
657 Node* u1 = NULL;
658 Node* u2 = NULL;
659 for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
660 Node* t1 = s1->fast_out(i);
661 if (!in_bb(t1)) continue;
662 for (DUIterator_Fast jmax, j = s2->fast_outs(jmax); j < jmax; j++) {
663 Node* t2 = s2->fast_out(j);
664 if (!in_bb(t2)) continue;
665 if (!opnd_positions_match(s1, t1, s2, t2))
666 continue;
667 if (stmts_can_pack(t1, t2, align)) {
668 int my_savings = est_savings(t1, t2);
669 if (my_savings > savings) {
670 savings = my_savings;
671 u1 = t1;
672 u2 = t2;
673 }
674 }
675 }
676 }
677 if (savings >= 0) {
678 Node_List* pair = new Node_List();
679 pair->push(u1);
680 pair->push(u2);
681 _packset.append(pair);
682 set_alignment(u1, u2, align);
683 changed = true;
684 }
685 return changed;
686 }
687
688 //---------------------------opnd_positions_match-------------------------
689 // Is the use of d1 in u1 at the same operand position as d2 in u2?
690 bool SuperWord::opnd_positions_match(Node* d1, Node* u1, Node* d2, Node* u2) {
691 uint ct = u1->req();
692 if (ct != u2->req()) return false;
693 uint i1 = 0;
694 uint i2 = 0;
695 do {
696 for (i1++; i1 < ct; i1++) if (u1->in(i1) == d1) break;
697 for (i2++; i2 < ct; i2++) if (u2->in(i2) == d2) break;
698 if (i1 != i2) {
699 return false;
700 }
701 } while (i1 < ct);
702 return true;
703 }
704
705 //------------------------------est_savings---------------------------
706 // Estimate the savings from executing s1 and s2 as a pack
707 int SuperWord::est_savings(Node* s1, Node* s2) {
708 int save = 2 - 1; // 2 operations per instruction in packed form
709
710 // inputs
711 for (uint i = 1; i < s1->req(); i++) {
712 Node* x1 = s1->in(i);
713 Node* x2 = s2->in(i);
714 if (x1 != x2) {
715 if (are_adjacent_refs(x1, x2)) {
716 save += adjacent_profit(x1, x2);
717 } else if (!in_packset(x1, x2)) {
718 save -= pack_cost(2);
719 } else {
720 save += unpack_cost(2);
721 }
722 }
723 }
724
725 // uses of result
726 uint ct = 0;
727 for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
728 Node* s1_use = s1->fast_out(i);
729 for (int j = 0; j < _packset.length(); j++) {
730 Node_List* p = _packset.at(j);
731 if (p->at(0) == s1_use) {
732 for (DUIterator_Fast kmax, k = s2->fast_outs(kmax); k < kmax; k++) {
733 Node* s2_use = s2->fast_out(k);
734 if (p->at(p->size()-1) == s2_use) {
735 ct++;
736 if (are_adjacent_refs(s1_use, s2_use)) {
737 save += adjacent_profit(s1_use, s2_use);
738 }
739 }
740 }
741 }
742 }
743 }
744
745 if (ct < s1->outcnt()) save += unpack_cost(1);
746 if (ct < s2->outcnt()) save += unpack_cost(1);
747
748 return save;
749 }
750
751 //------------------------------costs---------------------------
752 int SuperWord::adjacent_profit(Node* s1, Node* s2) { return 2; }
753 int SuperWord::pack_cost(int ct) { return ct; }
754 int SuperWord::unpack_cost(int ct) { return ct; }
755
756 //------------------------------combine_packs---------------------------
757 // Combine packs A and B with A.last == B.first into A.first..,A.last,B.second,..B.last
758 void SuperWord::combine_packs() {
759 bool changed;
760 do {
761 changed = false;
762 for (int i = 0; i < _packset.length(); i++) {
763 Node_List* p1 = _packset.at(i);
764 if (p1 == NULL) continue;
765 for (int j = 0; j < _packset.length(); j++) {
766 Node_List* p2 = _packset.at(j);
767 if (p2 == NULL) continue;
768 if (p1->at(p1->size()-1) == p2->at(0)) {
769 for (uint k = 1; k < p2->size(); k++) {
770 p1->push(p2->at(k));
771 }
772 _packset.at_put(j, NULL);
773 changed = true;
774 }
775 }
776 }
777 } while (changed);
778
779 for (int i = _packset.length() - 1; i >= 0; i--) {
780 Node_List* p1 = _packset.at(i);
781 if (p1 == NULL) {
782 _packset.remove_at(i);
783 }
784 }
785
786 #ifndef PRODUCT
787 if (TraceSuperWord) {
788 tty->print_cr("\nAfter combine_packs");
789 print_packset();
790 }
791 #endif
792 }
793
794 //-----------------------------construct_my_pack_map--------------------------
795 // Construct the map from nodes to packs. Only valid after the
796 // point where a node is only in one pack (after combine_packs).
797 void SuperWord::construct_my_pack_map() {
798 Node_List* rslt = NULL;
799 for (int i = 0; i < _packset.length(); i++) {
800 Node_List* p = _packset.at(i);
801 for (uint j = 0; j < p->size(); j++) {
802 Node* s = p->at(j);
803 assert(my_pack(s) == NULL, "only in one pack");
804 set_my_pack(s, p);
805 }
806 }
807 }
808
809 //------------------------------filter_packs---------------------------
810 // Remove packs that are not implemented or not profitable.
811 void SuperWord::filter_packs() {
812
813 // Remove packs that are not implemented
814 for (int i = _packset.length() - 1; i >= 0; i--) {
815 Node_List* pk = _packset.at(i);
816 bool impl = implemented(pk);
817 if (!impl) {
818 #ifndef PRODUCT
819 if (TraceSuperWord && Verbose) {
820 tty->print_cr("Unimplemented");
821 pk->at(0)->dump();
822 }
823 #endif
824 remove_pack_at(i);
825 }
826 }
827
828 // Remove packs that are not profitable
829 bool changed;
830 do {
831 changed = false;
832 for (int i = _packset.length() - 1; i >= 0; i--) {
833 Node_List* pk = _packset.at(i);
834 bool prof = profitable(pk);
835 if (!prof) {
836 #ifndef PRODUCT
837 if (TraceSuperWord && Verbose) {
838 tty->print_cr("Unprofitable");
839 pk->at(0)->dump();
840 }
841 #endif
842 remove_pack_at(i);
843 changed = true;
844 }
845 }
846 } while (changed);
847
848 #ifndef PRODUCT
849 if (TraceSuperWord) {
850 tty->print_cr("\nAfter filter_packs");
851 print_packset();
852 tty->cr();
853 }
854 #endif
855 }
856
857 //------------------------------implemented---------------------------
858 // Can code be generated for pack p?
859 bool SuperWord::implemented(Node_List* p) {
860 Node* p0 = p->at(0);
861 int vopc = VectorNode::opcode(p0->Opcode(), p->size(), velt_type(p0));
862 return vopc > 0 && Matcher::has_match_rule(vopc);
863 }
864
865 //------------------------------profitable---------------------------
866 // For pack p, are all operands and all uses (with in the block) vector?
867 bool SuperWord::profitable(Node_List* p) {
868 Node* p0 = p->at(0);
869 uint start, end;
870 vector_opd_range(p0, &start, &end);
871
872 // Return false if some input is not vector and inside block
873 for (uint i = start; i < end; i++) {
874 if (!is_vector_use(p0, i)) {
875 // For now, return false if not scalar promotion case (inputs are the same.)
876 // Later, implement PackNode and allow differing, non-vector inputs
877 // (maybe just the ones from outside the block.)
878 Node* p0_def = p0->in(i);
879 for (uint j = 1; j < p->size(); j++) {
880 Node* use = p->at(j);
881 Node* def = use->in(i);
882 if (p0_def != def)
883 return false;
884 }
885 }
886 }
887 if (!p0->is_Store()) {
888 // For now, return false if not all uses are vector.
889 // Later, implement ExtractNode and allow non-vector uses (maybe
890 // just the ones outside the block.)
891 for (uint i = 0; i < p->size(); i++) {
892 Node* def = p->at(i);
893 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
894 Node* use = def->fast_out(j);
895 for (uint k = 0; k < use->req(); k++) {
896 Node* n = use->in(k);
897 if (def == n) {
898 if (!is_vector_use(use, k)) {
899 return false;
900 }
901 }
902 }
903 }
904 }
905 }
906 return true;
907 }
908
909 //------------------------------schedule---------------------------
910 // Adjust the memory graph for the packed operations
911 void SuperWord::schedule() {
912
913 // Co-locate in the memory graph the members of each memory pack
914 for (int i = 0; i < _packset.length(); i++) {
915 co_locate_pack(_packset.at(i));
916 }
917 }
918
919 //-------------------------------remove_and_insert-------------------
920 //remove "current" from its current position in the memory graph and insert
921 //it after the appropriate insertion point (lip or uip)
922 void SuperWord::remove_and_insert(MemNode *current, MemNode *prev, MemNode *lip,
923 Node *uip, Unique_Node_List &sched_before) {
924 Node* my_mem = current->in(MemNode::Memory);
925 _igvn.hash_delete(current);
926 _igvn.hash_delete(my_mem);
927
928 //remove current_store from its current position in the memmory graph
929 for (DUIterator i = current->outs(); current->has_out(i); i++) {
930 Node* use = current->out(i);
931 if (use->is_Mem()) {
932 assert(use->in(MemNode::Memory) == current, "must be");
933 _igvn.hash_delete(use);
934 if (use == prev) { // connect prev to my_mem
935 use->set_req(MemNode::Memory, my_mem);
936 } else if (sched_before.member(use)) {
937 _igvn.hash_delete(uip);
938 use->set_req(MemNode::Memory, uip);
939 } else {
940 _igvn.hash_delete(lip);
941 use->set_req(MemNode::Memory, lip);
942 }
943 _igvn._worklist.push(use);
944 --i; //deleted this edge; rescan position
945 }
946 }
947
948 bool sched_up = sched_before.member(current);
949 Node *insert_pt = sched_up ? uip : lip;
950 _igvn.hash_delete(insert_pt);
951
952 // all uses of insert_pt's memory state should use current's instead
953 for (DUIterator i = insert_pt->outs(); insert_pt->has_out(i); i++) {
954 Node* use = insert_pt->out(i);
955 if (use->is_Mem()) {
956 assert(use->in(MemNode::Memory) == insert_pt, "must be");
957 _igvn.hash_delete(use);
958 use->set_req(MemNode::Memory, current);
959 _igvn._worklist.push(use);
960 --i; //deleted this edge; rescan position
961 } else if (!sched_up && use->is_Phi() && use->bottom_type() == Type::MEMORY) {
962 uint pos; //lip (lower insert point) must be the last one in the memory slice
963 _igvn.hash_delete(use);
964 for (pos=1; pos < use->req(); pos++) {
965 if (use->in(pos) == insert_pt) break;
966 }
967 use->set_req(pos, current);
968 _igvn._worklist.push(use);
969 --i;
970 }
971 }
972
973 //connect current to insert_pt
974 current->set_req(MemNode::Memory, insert_pt);
975 _igvn._worklist.push(current);
976 }
977
978 //------------------------------co_locate_pack----------------------------------
979 // To schedule a store pack, we need to move any sandwiched memory ops either before
980 // or after the pack, based upon dependence information:
981 // (1) If any store in the pack depends on the sandwiched memory op, the
982 // sandwiched memory op must be scheduled BEFORE the pack;
983 // (2) If a sandwiched memory op depends on any store in the pack, the
984 // sandwiched memory op must be scheduled AFTER the pack;
985 // (3) If a sandwiched memory op (say, memA) depends on another sandwiched
986 // memory op (say memB), memB must be scheduled before memA. So, if memA is
987 // scheduled before the pack, memB must also be scheduled before the pack;
988 // (4) If there is no dependence restriction for a sandwiched memory op, we simply
989 // schedule this store AFTER the pack
990 // (5) We know there is no dependence cycle, so there in no other case;
991 // (6) Finally, all memory ops in another single pack should be moved in the same direction.
992 //
993 // To schedule a load pack: the memory edge of every loads in the pack must be
994 // the same as the memory edge of the last executed load in the pack
995 void SuperWord::co_locate_pack(Node_List* pk) {
996 if (pk->at(0)->is_Store()) {
997 MemNode* first = executed_first(pk)->as_Mem();
998 MemNode* last = executed_last(pk)->as_Mem();
999 Unique_Node_List schedule_before_pack;
1000 Unique_Node_List memops;
1001
1002 MemNode* current = last->in(MemNode::Memory)->as_Mem();
1003 MemNode* previous = last;
1004 while (true) {
1005 assert(in_bb(current), "stay in block");
1006 memops.push(previous);
1007 for (DUIterator i = current->outs(); current->has_out(i); i++) {
1008 Node* use = current->out(i);
1009 if (use->is_Mem() && use != previous)
1010 memops.push(use);
1011 }
1012 if(current == first) break;
1013 previous = current;
1014 current = current->in(MemNode::Memory)->as_Mem();
1015 }
1016
1017 // determine which memory operations should be scheduled before the pack
1018 for (uint i = 1; i < memops.size(); i++) {
1019 Node *s1 = memops.at(i);
1020 if (!in_pack(s1, pk) && !schedule_before_pack.member(s1)) {
1021 for (uint j = 0; j< i; j++) {
1022 Node *s2 = memops.at(j);
1023 if (!independent(s1, s2)) {
1024 if (in_pack(s2, pk) || schedule_before_pack.member(s2)) {
1025 schedule_before_pack.push(s1); //s1 must be scheduled before
1026 Node_List* mem_pk = my_pack(s1);
1027 if (mem_pk != NULL) {
1028 for (uint ii = 0; ii < mem_pk->size(); ii++) {
1029 Node* s = mem_pk->at(ii); // follow partner
1030 if (memops.member(s) && !schedule_before_pack.member(s))
1031 schedule_before_pack.push(s);
1032 }
1033 }
1034 }
1035 }
1036 }
1037 }
1038 }
1039
1040 MemNode* lower_insert_pt = last;
1041 Node* upper_insert_pt = first->in(MemNode::Memory);
1042 previous = last; //previous store in pk
1043 current = last->in(MemNode::Memory)->as_Mem();
1044
1045 //start scheduling from "last" to "first"
1046 while (true) {
1047 assert(in_bb(current), "stay in block");
1048 assert(in_pack(previous, pk), "previous stays in pack");
1049 Node* my_mem = current->in(MemNode::Memory);
1050
1051 if (in_pack(current, pk)) {
1052 // Forward users of my memory state (except "previous) to my input memory state
1053 _igvn.hash_delete(current);
1054 for (DUIterator i = current->outs(); current->has_out(i); i++) {
1055 Node* use = current->out(i);
1056 if (use->is_Mem() && use != previous) {
1057 assert(use->in(MemNode::Memory) == current, "must be");
1058 _igvn.hash_delete(use);
1059 if (schedule_before_pack.member(use)) {
1060 _igvn.hash_delete(upper_insert_pt);
1061 use->set_req(MemNode::Memory, upper_insert_pt);
1062 } else {
1063 _igvn.hash_delete(lower_insert_pt);
1064 use->set_req(MemNode::Memory, lower_insert_pt);
1065 }
1066 _igvn._worklist.push(use);
1067 --i; // deleted this edge; rescan position
1068 }
1069 }
1070 previous = current;
1071 } else { // !in_pack(current, pk) ==> a sandwiched store
1072 remove_and_insert(current, previous, lower_insert_pt, upper_insert_pt, schedule_before_pack);
1073 }
1074
1075 if (current == first) break;
1076 current = my_mem->as_Mem();
1077 } // end while
1078 } else if (pk->at(0)->is_Load()) { //load
1079 // all use the memory state that the last executed load uses
1080 LoadNode* last_load = executed_last(pk)->as_Load();
1081 Node* last_mem = last_load->in(MemNode::Memory);
1082 _igvn.hash_delete(last_mem);
1083 // Give each load same memory state as last
1084 for (uint i = 0; i < pk->size(); i++) {
1085 LoadNode* ld = pk->at(i)->as_Load();
1086 _igvn.hash_delete(ld);
1087 ld->set_req(MemNode::Memory, last_mem);
1088 _igvn._worklist.push(ld);
1089 }
1090 }
1091 }
1092
1093 //------------------------------output---------------------------
1094 // Convert packs into vector node operations
1095 void SuperWord::output() {
1096 if (_packset.length() == 0) return;
1097
1098 // MUST ENSURE main loop's initial value is properly aligned:
1099 // (iv_initial_value + min_iv_offset) % vector_width_in_bytes() == 0
1100
1101 align_initial_loop_index(align_to_ref());
1102
1103 // Insert extract (unpack) operations for scalar uses
1104 for (int i = 0; i < _packset.length(); i++) {
1105 insert_extracts(_packset.at(i));
1106 }
1107
1108 for (int i = 0; i < _block.length(); i++) {
1109 Node* n = _block.at(i);
1110 Node_List* p = my_pack(n);
1111 if (p && n == executed_last(p)) {
1112 uint vlen = p->size();
1113 Node* vn = NULL;
1114 Node* low_adr = p->at(0);
1115 Node* first = executed_first(p);
1116 if (n->is_Load()) {
1117 int opc = n->Opcode();
1118 Node* ctl = n->in(MemNode::Control);
1119 Node* mem = first->in(MemNode::Memory);
1120 Node* adr = low_adr->in(MemNode::Address);
1121 const TypePtr* atyp = n->adr_type();
1122 vn = VectorLoadNode::make(_phase->C, opc, ctl, mem, adr, atyp, vlen);
1123
1124 } else if (n->is_Store()) {
1125 // Promote value to be stored to vector
1126 VectorNode* val = vector_opd(p, MemNode::ValueIn);
1127
1128 int opc = n->Opcode();
1129 Node* ctl = n->in(MemNode::Control);
1130 Node* mem = first->in(MemNode::Memory);
1131 Node* adr = low_adr->in(MemNode::Address);
1132 const TypePtr* atyp = n->adr_type();
1133 vn = VectorStoreNode::make(_phase->C, opc, ctl, mem, adr, atyp, val, vlen);
1134
1135 } else if (n->req() == 3) {
1136 // Promote operands to vector
1137 Node* in1 = vector_opd(p, 1);
1138 Node* in2 = vector_opd(p, 2);
1139 vn = VectorNode::make(_phase->C, n->Opcode(), in1, in2, vlen, velt_type(n));
1140
1141 } else {
1142 ShouldNotReachHere();
1143 }
1144
1145 _phase->_igvn.register_new_node_with_optimizer(vn);
1146 _phase->set_ctrl(vn, _phase->get_ctrl(p->at(0)));
1147 for (uint j = 0; j < p->size(); j++) {
1148 Node* pm = p->at(j);
1149 _igvn.hash_delete(pm);
1150 _igvn.subsume_node(pm, vn);
1151 }
1152 _igvn._worklist.push(vn);
1153 }
1154 }
1155 }
1156
1157 //------------------------------vector_opd---------------------------
1158 // Create a vector operand for the nodes in pack p for operand: in(opd_idx)
1159 VectorNode* SuperWord::vector_opd(Node_List* p, int opd_idx) {
1160 Node* p0 = p->at(0);
1161 uint vlen = p->size();
1162 Node* opd = p0->in(opd_idx);
1163
1164 bool same_opd = true;
1165 for (uint i = 1; i < vlen; i++) {
1166 Node* pi = p->at(i);
1167 Node* in = pi->in(opd_idx);
1168 if (opd != in) {
1169 same_opd = false;
1170 break;
1171 }
1172 }
1173
1174 if (same_opd) {
1175 if (opd->is_Vector()) {
1176 return (VectorNode*)opd; // input is matching vector
1177 }
1178 // Convert scalar input to vector. Use p0's type because it's container
1179 // maybe smaller than the operand's container.
1180 const Type* opd_t = velt_type(!in_bb(opd) ? p0 : opd);
1181 const Type* p0_t = velt_type(p0);
1182 if (p0_t->higher_equal(opd_t)) opd_t = p0_t;
1183 VectorNode* vn = VectorNode::scalar2vector(_phase->C, opd, vlen, opd_t);
1184
1185 _phase->_igvn.register_new_node_with_optimizer(vn);
1186 _phase->set_ctrl(vn, _phase->get_ctrl(opd));
1187 return vn;
1188 }
1189
1190 // Insert pack operation
1191 const Type* opd_t = velt_type(!in_bb(opd) ? p0 : opd);
1192 PackNode* pk = PackNode::make(_phase->C, opd, opd_t);
1193
1194 for (uint i = 1; i < vlen; i++) {
1195 Node* pi = p->at(i);
1196 Node* in = pi->in(opd_idx);
1197 assert(my_pack(in) == NULL, "Should already have been unpacked");
1198 assert(opd_t == velt_type(!in_bb(in) ? pi : in), "all same type");
1199 pk->add_opd(in);
1200 }
1201 _phase->_igvn.register_new_node_with_optimizer(pk);
1202 _phase->set_ctrl(pk, _phase->get_ctrl(opd));
1203 return pk;
1204 }
1205
1206 //------------------------------insert_extracts---------------------------
1207 // If a use of pack p is not a vector use, then replace the
1208 // use with an extract operation.
1209 void SuperWord::insert_extracts(Node_List* p) {
1210 if (p->at(0)->is_Store()) return;
1211 assert(_n_idx_list.is_empty(), "empty (node,index) list");
1212
1213 // Inspect each use of each pack member. For each use that is
1214 // not a vector use, replace the use with an extract operation.
1215
1216 for (uint i = 0; i < p->size(); i++) {
1217 Node* def = p->at(i);
1218 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
1219 Node* use = def->fast_out(j);
1220 for (uint k = 0; k < use->req(); k++) {
1221 Node* n = use->in(k);
1222 if (def == n) {
1223 if (!is_vector_use(use, k)) {
1224 _n_idx_list.push(use, k);
1225 }
1226 }
1227 }
1228 }
1229 }
1230
1231 while (_n_idx_list.is_nonempty()) {
1232 Node* use = _n_idx_list.node();
1233 int idx = _n_idx_list.index();
1234 _n_idx_list.pop();
1235 Node* def = use->in(idx);
1236
1237 // Insert extract operation
1238 _igvn.hash_delete(def);
1239 _igvn.hash_delete(use);
1240 int def_pos = alignment(def) / data_size(def);
1241 const Type* def_t = velt_type(def);
1242
1243 Node* ex = ExtractNode::make(_phase->C, def, def_pos, def_t);
1244 _phase->_igvn.register_new_node_with_optimizer(ex);
1245 _phase->set_ctrl(ex, _phase->get_ctrl(def));
1246 use->set_req(idx, ex);
1247 _igvn._worklist.push(def);
1248 _igvn._worklist.push(use);
1249
1250 bb_insert_after(ex, bb_idx(def));
1251 set_velt_type(ex, def_t);
1252 }
1253 }
1254
1255 //------------------------------is_vector_use---------------------------
1256 // Is use->in(u_idx) a vector use?
1257 bool SuperWord::is_vector_use(Node* use, int u_idx) {
1258 Node_List* u_pk = my_pack(use);
1259 if (u_pk == NULL) return false;
1260 Node* def = use->in(u_idx);
1261 Node_List* d_pk = my_pack(def);
1262 if (d_pk == NULL) {
1263 // check for scalar promotion
1264 Node* n = u_pk->at(0)->in(u_idx);
1265 for (uint i = 1; i < u_pk->size(); i++) {
1266 if (u_pk->at(i)->in(u_idx) != n) return false;
1267 }
1268 return true;
1269 }
1270 if (u_pk->size() != d_pk->size())
1271 return false;
1272 for (uint i = 0; i < u_pk->size(); i++) {
1273 Node* ui = u_pk->at(i);
1274 Node* di = d_pk->at(i);
1275 if (ui->in(u_idx) != di || alignment(ui) != alignment(di))
1276 return false;
1277 }
1278 return true;
1279 }
1280
1281 //------------------------------construct_bb---------------------------
1282 // Construct reverse postorder list of block members
1283 void SuperWord::construct_bb() {
1284 Node* entry = bb();
1285
1286 assert(_stk.length() == 0, "stk is empty");
1287 assert(_block.length() == 0, "block is empty");
1288 assert(_data_entry.length() == 0, "data_entry is empty");
1289 assert(_mem_slice_head.length() == 0, "mem_slice_head is empty");
1290 assert(_mem_slice_tail.length() == 0, "mem_slice_tail is empty");
1291
1292 // Find non-control nodes with no inputs from within block,
1293 // create a temporary map from node _idx to bb_idx for use
1294 // by the visited and post_visited sets,
1295 // and count number of nodes in block.
1296 int bb_ct = 0;
1297 for (uint i = 0; i < lpt()->_body.size(); i++ ) {
1298 Node *n = lpt()->_body.at(i);
1299 set_bb_idx(n, i); // Create a temporary map
1300 if (in_bb(n)) {
1301 bb_ct++;
1302 if (!n->is_CFG()) {
1303 bool found = false;
1304 for (uint j = 0; j < n->req(); j++) {
1305 Node* def = n->in(j);
1306 if (def && in_bb(def)) {
1307 found = true;
1308 break;
1309 }
1310 }
1311 if (!found) {
1312 assert(n != entry, "can't be entry");
1313 _data_entry.push(n);
1314 }
1315 }
1316 }
1317 }
1318
1319 // Find memory slices (head and tail)
1320 for (DUIterator_Fast imax, i = lp()->fast_outs(imax); i < imax; i++) {
1321 Node *n = lp()->fast_out(i);
1322 if (in_bb(n) && (n->is_Phi() && n->bottom_type() == Type::MEMORY)) {
1323 Node* n_tail = n->in(LoopNode::LoopBackControl);
1324 if (n_tail != n->in(LoopNode::EntryControl)) {
1325 _mem_slice_head.push(n);
1326 _mem_slice_tail.push(n_tail);
1327 }
1328 }
1329 }
1330
1331 // Create an RPO list of nodes in block
1332
1333 visited_clear();
1334 post_visited_clear();
1335
1336 // Push all non-control nodes with no inputs from within block, then control entry
1337 for (int j = 0; j < _data_entry.length(); j++) {
1338 Node* n = _data_entry.at(j);
1339 visited_set(n);
1340 _stk.push(n);
1341 }
1342 visited_set(entry);
1343 _stk.push(entry);
1344
1345 // Do a depth first walk over out edges
1346 int rpo_idx = bb_ct - 1;
1347 int size;
1348 while ((size = _stk.length()) > 0) {
1349 Node* n = _stk.top(); // Leave node on stack
1350 if (!visited_test_set(n)) {
1351 // forward arc in graph
1352 } else if (!post_visited_test(n)) {
1353 // cross or back arc
1354 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1355 Node *use = n->fast_out(i);
1356 if (in_bb(use) && !visited_test(use) &&
1357 // Don't go around backedge
1358 (!use->is_Phi() || n == entry)) {
1359 _stk.push(use);
1360 }
1361 }
1362 if (_stk.length() == size) {
1363 // There were no additional uses, post visit node now
1364 _stk.pop(); // Remove node from stack
1365 assert(rpo_idx >= 0, "");
1366 _block.at_put_grow(rpo_idx, n);
1367 rpo_idx--;
1368 post_visited_set(n);
1369 assert(rpo_idx >= 0 || _stk.is_empty(), "");
1370 }
1371 } else {
1372 _stk.pop(); // Remove post-visited node from stack
1373 }
1374 }
1375
1376 // Create real map of block indices for nodes
1377 for (int j = 0; j < _block.length(); j++) {
1378 Node* n = _block.at(j);
1379 set_bb_idx(n, j);
1380 }
1381
1382 initialize_bb(); // Ensure extra info is allocated.
1383
1384 #ifndef PRODUCT
1385 if (TraceSuperWord) {
1386 print_bb();
1387 tty->print_cr("\ndata entry nodes: %s", _data_entry.length() > 0 ? "" : "NONE");
1388 for (int m = 0; m < _data_entry.length(); m++) {
1389 tty->print("%3d ", m);
1390 _data_entry.at(m)->dump();
1391 }
1392 tty->print_cr("\nmemory slices: %s", _mem_slice_head.length() > 0 ? "" : "NONE");
1393 for (int m = 0; m < _mem_slice_head.length(); m++) {
1394 tty->print("%3d ", m); _mem_slice_head.at(m)->dump();
1395 tty->print(" "); _mem_slice_tail.at(m)->dump();
1396 }
1397 }
1398 #endif
1399 assert(rpo_idx == -1 && bb_ct == _block.length(), "all block members found");
1400 }
1401
1402 //------------------------------initialize_bb---------------------------
1403 // Initialize per node info
1404 void SuperWord::initialize_bb() {
1405 Node* last = _block.at(_block.length() - 1);
1406 grow_node_info(bb_idx(last));
1407 }
1408
1409 //------------------------------bb_insert_after---------------------------
1410 // Insert n into block after pos
1411 void SuperWord::bb_insert_after(Node* n, int pos) {
1412 int n_pos = pos + 1;
1413 // Make room
1414 for (int i = _block.length() - 1; i >= n_pos; i--) {
1415 _block.at_put_grow(i+1, _block.at(i));
1416 }
1417 for (int j = _node_info.length() - 1; j >= n_pos; j--) {
1418 _node_info.at_put_grow(j+1, _node_info.at(j));
1419 }
1420 // Set value
1421 _block.at_put_grow(n_pos, n);
1422 _node_info.at_put_grow(n_pos, SWNodeInfo::initial);
1423 // Adjust map from node->_idx to _block index
1424 for (int i = n_pos; i < _block.length(); i++) {
1425 set_bb_idx(_block.at(i), i);
1426 }
1427 }
1428
1429 //------------------------------compute_max_depth---------------------------
1430 // Compute max depth for expressions from beginning of block
1431 // Use to prune search paths during test for independence.
1432 void SuperWord::compute_max_depth() {
1433 int ct = 0;
1434 bool again;
1435 do {
1436 again = false;
1437 for (int i = 0; i < _block.length(); i++) {
1438 Node* n = _block.at(i);
1439 if (!n->is_Phi()) {
1440 int d_orig = depth(n);
1441 int d_in = 0;
1442 for (DepPreds preds(n, _dg); !preds.done(); preds.next()) {
1443 Node* pred = preds.current();
1444 if (in_bb(pred)) {
1445 d_in = MAX2(d_in, depth(pred));
1446 }
1447 }
1448 if (d_in + 1 != d_orig) {
1449 set_depth(n, d_in + 1);
1450 again = true;
1451 }
1452 }
1453 }
1454 ct++;
1455 } while (again);
1456 #ifndef PRODUCT
1457 if (TraceSuperWord && Verbose)
1458 tty->print_cr("compute_max_depth iterated: %d times", ct);
1459 #endif
1460 }
1461
1462 //-------------------------compute_vector_element_type-----------------------
1463 // Compute necessary vector element type for expressions
1464 // This propagates backwards a narrower integer type when the
1465 // upper bits of the value are not needed.
1466 // Example: char a,b,c; a = b + c;
1467 // Normally the type of the add is integer, but for packed character
1468 // operations the type of the add needs to be char.
1469 void SuperWord::compute_vector_element_type() {
1470 #ifndef PRODUCT
1471 if (TraceSuperWord && Verbose)
1472 tty->print_cr("\ncompute_velt_type:");
1473 #endif
1474
1475 // Initial type
1476 for (int i = 0; i < _block.length(); i++) {
1477 Node* n = _block.at(i);
1478 const Type* t = n->is_Mem() ? Type::get_const_basic_type(n->as_Mem()->memory_type())
1479 : _igvn.type(n);
1480 const Type* vt = container_type(t);
1481 set_velt_type(n, vt);
1482 }
1483
1484 // Propagate narrowed type backwards through operations
1485 // that don't depend on higher order bits
1486 for (int i = _block.length() - 1; i >= 0; i--) {
1487 Node* n = _block.at(i);
1488 // Only integer types need be examined
1489 if (n->bottom_type()->isa_int()) {
1490 uint start, end;
1491 vector_opd_range(n, &start, &end);
1492 const Type* vt = velt_type(n);
1493
1494 for (uint j = start; j < end; j++) {
1495 Node* in = n->in(j);
1496 // Don't propagate through a type conversion
1497 if (n->bottom_type() != in->bottom_type())
1498 continue;
1499 switch(in->Opcode()) {
1500 case Op_AddI: case Op_AddL:
1501 case Op_SubI: case Op_SubL:
1502 case Op_MulI: case Op_MulL:
1503 case Op_AndI: case Op_AndL:
1504 case Op_OrI: case Op_OrL:
1505 case Op_XorI: case Op_XorL:
1506 case Op_LShiftI: case Op_LShiftL:
1507 case Op_CMoveI: case Op_CMoveL:
1508 if (in_bb(in)) {
1509 bool same_type = true;
1510 for (DUIterator_Fast kmax, k = in->fast_outs(kmax); k < kmax; k++) {
1511 Node *use = in->fast_out(k);
1512 if (!in_bb(use) || velt_type(use) != vt) {
1513 same_type = false;
1514 break;
1515 }
1516 }
1517 if (same_type) {
1518 set_velt_type(in, vt);
1519 }
1520 }
1521 }
1522 }
1523 }
1524 }
1525 #ifndef PRODUCT
1526 if (TraceSuperWord && Verbose) {
1527 for (int i = 0; i < _block.length(); i++) {
1528 Node* n = _block.at(i);
1529 velt_type(n)->dump();
1530 tty->print("\t");
1531 n->dump();
1532 }
1533 }
1534 #endif
1535 }
1536
1537 //------------------------------memory_alignment---------------------------
1538 // Alignment within a vector memory reference
1539 int SuperWord::memory_alignment(MemNode* s, int iv_adjust_in_bytes) {
1540 SWPointer p(s, this);
1541 if (!p.valid()) {
1542 return bottom_align;
1543 }
1544 int offset = p.offset_in_bytes();
1545 offset += iv_adjust_in_bytes;
1546 int off_rem = offset % vector_width_in_bytes();
1547 int off_mod = off_rem >= 0 ? off_rem : off_rem + vector_width_in_bytes();
1548 return off_mod;
1549 }
1550
1551 //---------------------------container_type---------------------------
1552 // Smallest type containing range of values
1553 const Type* SuperWord::container_type(const Type* t) {
1554 const Type* tp = t->make_ptr();
1555 if (tp && tp->isa_aryptr()) {
1556 t = tp->is_aryptr()->elem();
1557 }
1558 if (t->basic_type() == T_INT) {
1559 if (t->higher_equal(TypeInt::BOOL)) return TypeInt::BOOL;
1560 if (t->higher_equal(TypeInt::BYTE)) return TypeInt::BYTE;
1561 if (t->higher_equal(TypeInt::CHAR)) return TypeInt::CHAR;
1562 if (t->higher_equal(TypeInt::SHORT)) return TypeInt::SHORT;
1563 return TypeInt::INT;
1564 }
1565 return t;
1566 }
1567
1568 //-------------------------vector_opd_range-----------------------
1569 // (Start, end] half-open range defining which operands are vector
1570 void SuperWord::vector_opd_range(Node* n, uint* start, uint* end) {
1571 switch (n->Opcode()) {
1572 case Op_LoadB: case Op_LoadUS:
1573 case Op_LoadI: case Op_LoadL:
1574 case Op_LoadF: case Op_LoadD:
1575 case Op_LoadP:
1576 *start = 0;
1577 *end = 0;
1578 return;
1579 case Op_StoreB: case Op_StoreC:
1580 case Op_StoreI: case Op_StoreL:
1581 case Op_StoreF: case Op_StoreD:
1582 case Op_StoreP:
1583 *start = MemNode::ValueIn;
1584 *end = *start + 1;
1585 return;
1586 case Op_LShiftI: case Op_LShiftL:
1587 *start = 1;
1588 *end = 2;
1589 return;
1590 case Op_CMoveI: case Op_CMoveL: case Op_CMoveF: case Op_CMoveD:
1591 *start = 2;
1592 *end = n->req();
1593 return;
1594 }
1595 *start = 1;
1596 *end = n->req(); // default is all operands
1597 }
1598
1599 //------------------------------in_packset---------------------------
1600 // Are s1 and s2 in a pack pair and ordered as s1,s2?
1601 bool SuperWord::in_packset(Node* s1, Node* s2) {
1602 for (int i = 0; i < _packset.length(); i++) {
1603 Node_List* p = _packset.at(i);
1604 assert(p->size() == 2, "must be");
1605 if (p->at(0) == s1 && p->at(p->size()-1) == s2) {
1606 return true;
1607 }
1608 }
1609 return false;
1610 }
1611
1612 //------------------------------in_pack---------------------------
1613 // Is s in pack p?
1614 Node_List* SuperWord::in_pack(Node* s, Node_List* p) {
1615 for (uint i = 0; i < p->size(); i++) {
1616 if (p->at(i) == s) {
1617 return p;
1618 }
1619 }
1620 return NULL;
1621 }
1622
1623 //------------------------------remove_pack_at---------------------------
1624 // Remove the pack at position pos in the packset
1625 void SuperWord::remove_pack_at(int pos) {
1626 Node_List* p = _packset.at(pos);
1627 for (uint i = 0; i < p->size(); i++) {
1628 Node* s = p->at(i);
1629 set_my_pack(s, NULL);
1630 }
1631 _packset.remove_at(pos);
1632 }
1633
1634 //------------------------------executed_first---------------------------
1635 // Return the node executed first in pack p. Uses the RPO block list
1636 // to determine order.
1637 Node* SuperWord::executed_first(Node_List* p) {
1638 Node* n = p->at(0);
1639 int n_rpo = bb_idx(n);
1640 for (uint i = 1; i < p->size(); i++) {
1641 Node* s = p->at(i);
1642 int s_rpo = bb_idx(s);
1643 if (s_rpo < n_rpo) {
1644 n = s;
1645 n_rpo = s_rpo;
1646 }
1647 }
1648 return n;
1649 }
1650
1651 //------------------------------executed_last---------------------------
1652 // Return the node executed last in pack p.
1653 Node* SuperWord::executed_last(Node_List* p) {
1654 Node* n = p->at(0);
1655 int n_rpo = bb_idx(n);
1656 for (uint i = 1; i < p->size(); i++) {
1657 Node* s = p->at(i);
1658 int s_rpo = bb_idx(s);
1659 if (s_rpo > n_rpo) {
1660 n = s;
1661 n_rpo = s_rpo;
1662 }
1663 }
1664 return n;
1665 }
1666
1667 //----------------------------align_initial_loop_index---------------------------
1668 // Adjust pre-loop limit so that in main loop, a load/store reference
1669 // to align_to_ref will be a position zero in the vector.
1670 // (iv + k) mod vector_align == 0
1671 void SuperWord::align_initial_loop_index(MemNode* align_to_ref) {
1672 CountedLoopNode *main_head = lp()->as_CountedLoop();
1673 assert(main_head->is_main_loop(), "");
1674 CountedLoopEndNode* pre_end = get_pre_loop_end(main_head);
1675 assert(pre_end != NULL, "");
1676 Node *pre_opaq1 = pre_end->limit();
1677 assert(pre_opaq1->Opcode() == Op_Opaque1, "");
1678 Opaque1Node *pre_opaq = (Opaque1Node*)pre_opaq1;
1679 Node *lim0 = pre_opaq->in(1);
1680
1681 // Where we put new limit calculations
1682 Node *pre_ctrl = pre_end->loopnode()->in(LoopNode::EntryControl);
1683
1684 // Ensure the original loop limit is available from the
1685 // pre-loop Opaque1 node.
1686 Node *orig_limit = pre_opaq->original_loop_limit();
1687 assert(orig_limit != NULL && _igvn.type(orig_limit) != Type::TOP, "");
1688
1689 SWPointer align_to_ref_p(align_to_ref, this);
1690
1691 // Given:
1692 // lim0 == original pre loop limit
1693 // V == v_align (power of 2)
1694 // invar == extra invariant piece of the address expression
1695 // e == k [ +/- invar ]
1696 //
1697 // When reassociating expressions involving '%' the basic rules are:
1698 // (a - b) % k == 0 => a % k == b % k
1699 // and:
1700 // (a + b) % k == 0 => a % k == (k - b) % k
1701 //
1702 // For stride > 0 && scale > 0,
1703 // Derive the new pre-loop limit "lim" such that the two constraints:
1704 // (1) lim = lim0 + N (where N is some positive integer < V)
1705 // (2) (e + lim) % V == 0
1706 // are true.
1707 //
1708 // Substituting (1) into (2),
1709 // (e + lim0 + N) % V == 0
1710 // solve for N:
1711 // N = (V - (e + lim0)) % V
1712 // substitute back into (1), so that new limit
1713 // lim = lim0 + (V - (e + lim0)) % V
1714 //
1715 // For stride > 0 && scale < 0
1716 // Constraints:
1717 // lim = lim0 + N
1718 // (e - lim) % V == 0
1719 // Solving for lim:
1720 // (e - lim0 - N) % V == 0
1721 // N = (e - lim0) % V
1722 // lim = lim0 + (e - lim0) % V
1723 //
1724 // For stride < 0 && scale > 0
1725 // Constraints:
1726 // lim = lim0 - N
1727 // (e + lim) % V == 0
1728 // Solving for lim:
1729 // (e + lim0 - N) % V == 0
1730 // N = (e + lim0) % V
1731 // lim = lim0 - (e + lim0) % V
1732 //
1733 // For stride < 0 && scale < 0
1734 // Constraints:
1735 // lim = lim0 - N
1736 // (e - lim) % V == 0
1737 // Solving for lim:
1738 // (e - lim0 + N) % V == 0
1739 // N = (V - (e - lim0)) % V
1740 // lim = lim0 - (V - (e - lim0)) % V
1741
1742 int stride = iv_stride();
1743 int scale = align_to_ref_p.scale_in_bytes();
1744 int elt_size = align_to_ref_p.memory_size();
1745 int v_align = vector_width_in_bytes() / elt_size;
1746 int k = align_to_ref_p.offset_in_bytes() / elt_size;
1747
1748 Node *kn = _igvn.intcon(k);
1749
1750 Node *e = kn;
1751 if (align_to_ref_p.invar() != NULL) {
1752 // incorporate any extra invariant piece producing k +/- invar >>> log2(elt)
1753 Node* log2_elt = _igvn.intcon(exact_log2(elt_size));
1754 Node* aref = new (_phase->C, 3) URShiftINode(align_to_ref_p.invar(), log2_elt);
1755 _phase->_igvn.register_new_node_with_optimizer(aref);
1756 _phase->set_ctrl(aref, pre_ctrl);
1757 if (align_to_ref_p.negate_invar()) {
1758 e = new (_phase->C, 3) SubINode(e, aref);
1759 } else {
1760 e = new (_phase->C, 3) AddINode(e, aref);
1761 }
1762 _phase->_igvn.register_new_node_with_optimizer(e);
1763 _phase->set_ctrl(e, pre_ctrl);
1764 }
1765
1766 // compute e +/- lim0
1767 if (scale < 0) {
1768 e = new (_phase->C, 3) SubINode(e, lim0);
1769 } else {
1770 e = new (_phase->C, 3) AddINode(e, lim0);
1771 }
1772 _phase->_igvn.register_new_node_with_optimizer(e);
1773 _phase->set_ctrl(e, pre_ctrl);
1774
1775 if (stride * scale > 0) {
1776 // compute V - (e +/- lim0)
1777 Node* va = _igvn.intcon(v_align);
1778 e = new (_phase->C, 3) SubINode(va, e);
1779 _phase->_igvn.register_new_node_with_optimizer(e);
1780 _phase->set_ctrl(e, pre_ctrl);
1781 }
1782 // compute N = (exp) % V
1783 Node* va_msk = _igvn.intcon(v_align - 1);
1784 Node* N = new (_phase->C, 3) AndINode(e, va_msk);
1785 _phase->_igvn.register_new_node_with_optimizer(N);
1786 _phase->set_ctrl(N, pre_ctrl);
1787
1788 // substitute back into (1), so that new limit
1789 // lim = lim0 + N
1790 Node* lim;
1791 if (stride < 0) {
1792 lim = new (_phase->C, 3) SubINode(lim0, N);
1793 } else {
1794 lim = new (_phase->C, 3) AddINode(lim0, N);
1795 }
1796 _phase->_igvn.register_new_node_with_optimizer(lim);
1797 _phase->set_ctrl(lim, pre_ctrl);
1798 Node* constrained =
1799 (stride > 0) ? (Node*) new (_phase->C,3) MinINode(lim, orig_limit)
1800 : (Node*) new (_phase->C,3) MaxINode(lim, orig_limit);
1801 _phase->_igvn.register_new_node_with_optimizer(constrained);
1802 _phase->set_ctrl(constrained, pre_ctrl);
1803 _igvn.hash_delete(pre_opaq);
1804 pre_opaq->set_req(1, constrained);
1805 }
1806
1807 //----------------------------get_pre_loop_end---------------------------
1808 // Find pre loop end from main loop. Returns null if none.
1809 CountedLoopEndNode* SuperWord::get_pre_loop_end(CountedLoopNode *cl) {
1810 Node *ctrl = cl->in(LoopNode::EntryControl);
1811 if (!ctrl->is_IfTrue() && !ctrl->is_IfFalse()) return NULL;
1812 Node *iffm = ctrl->in(0);
1813 if (!iffm->is_If()) return NULL;
1814 Node *p_f = iffm->in(0);
1815 if (!p_f->is_IfFalse()) return NULL;
1816 if (!p_f->in(0)->is_CountedLoopEnd()) return NULL;
1817 CountedLoopEndNode *pre_end = p_f->in(0)->as_CountedLoopEnd();
1818 if (!pre_end->loopnode()->is_pre_loop()) return NULL;
1819 return pre_end;
1820 }
1821
1822
1823 //------------------------------init---------------------------
1824 void SuperWord::init() {
1825 _dg.init();
1826 _packset.clear();
1827 _disjoint_ptrs.clear();
1828 _block.clear();
1829 _data_entry.clear();
1830 _mem_slice_head.clear();
1831 _mem_slice_tail.clear();
1832 _node_info.clear();
1833 _align_to_ref = NULL;
1834 _lpt = NULL;
1835 _lp = NULL;
1836 _bb = NULL;
1837 _iv = NULL;
1838 }
1839
1840 //------------------------------print_packset---------------------------
1841 void SuperWord::print_packset() {
1842 #ifndef PRODUCT
1843 tty->print_cr("packset");
1844 for (int i = 0; i < _packset.length(); i++) {
1845 tty->print_cr("Pack: %d", i);
1846 Node_List* p = _packset.at(i);
1847 print_pack(p);
1848 }
1849 #endif
1850 }
1851
1852 //------------------------------print_pack---------------------------
1853 void SuperWord::print_pack(Node_List* p) {
1854 for (uint i = 0; i < p->size(); i++) {
1855 print_stmt(p->at(i));
1856 }
1857 }
1858
1859 //------------------------------print_bb---------------------------
1860 void SuperWord::print_bb() {
1861 #ifndef PRODUCT
1862 tty->print_cr("\nBlock");
1863 for (int i = 0; i < _block.length(); i++) {
1864 Node* n = _block.at(i);
1865 tty->print("%d ", i);
1866 if (n) {
1867 n->dump();
1868 }
1869 }
1870 #endif
1871 }
1872
1873 //------------------------------print_stmt---------------------------
1874 void SuperWord::print_stmt(Node* s) {
1875 #ifndef PRODUCT
1876 tty->print(" align: %d \t", alignment(s));
1877 s->dump();
1878 #endif
1879 }
1880
1881 //------------------------------blank---------------------------
1882 char* SuperWord::blank(uint depth) {
1883 static char blanks[101];
1884 assert(depth < 101, "too deep");
1885 for (uint i = 0; i < depth; i++) blanks[i] = ' ';
1886 blanks[depth] = '\0';
1887 return blanks;
1888 }
1889
1890
1891 //==============================SWPointer===========================
1892
1893 //----------------------------SWPointer------------------------
1894 SWPointer::SWPointer(MemNode* mem, SuperWord* slp) :
1895 _mem(mem), _slp(slp), _base(NULL), _adr(NULL),
1896 _scale(0), _offset(0), _invar(NULL), _negate_invar(false) {
1897
1898 Node* adr = mem->in(MemNode::Address);
1899 if (!adr->is_AddP()) {
1900 assert(!valid(), "too complex");
1901 return;
1902 }
1903 // Match AddP(base, AddP(ptr, k*iv [+ invariant]), constant)
1904 Node* base = adr->in(AddPNode::Base);
1905 for (int i = 0; i < 3; i++) {
1906 if (!scaled_iv_plus_offset(adr->in(AddPNode::Offset))) {
1907 assert(!valid(), "too complex");
1908 return;
1909 }
1910 adr = adr->in(AddPNode::Address);
1911 if (base == adr || !adr->is_AddP()) {
1912 break; // stop looking at addp's
1913 }
1914 }
1915 _base = base;
1916 _adr = adr;
1917 assert(valid(), "Usable");
1918 }
1919
1920 // Following is used to create a temporary object during
1921 // the pattern match of an address expression.
1922 SWPointer::SWPointer(SWPointer* p) :
1923 _mem(p->_mem), _slp(p->_slp), _base(NULL), _adr(NULL),
1924 _scale(0), _offset(0), _invar(NULL), _negate_invar(false) {}
1925
1926 //------------------------scaled_iv_plus_offset--------------------
1927 // Match: k*iv + offset
1928 // where: k is a constant that maybe zero, and
1929 // offset is (k2 [+/- invariant]) where k2 maybe zero and invariant is optional
1930 bool SWPointer::scaled_iv_plus_offset(Node* n) {
1931 if (scaled_iv(n)) {
1932 return true;
1933 }
1934 if (offset_plus_k(n)) {
1935 return true;
1936 }
1937 int opc = n->Opcode();
1938 if (opc == Op_AddI) {
1939 if (scaled_iv(n->in(1)) && offset_plus_k(n->in(2))) {
1940 return true;
1941 }
1942 if (scaled_iv(n->in(2)) && offset_plus_k(n->in(1))) {
1943 return true;
1944 }
1945 } else if (opc == Op_SubI) {
1946 if (scaled_iv(n->in(1)) && offset_plus_k(n->in(2), true)) {
1947 return true;
1948 }
1949 if (scaled_iv(n->in(2)) && offset_plus_k(n->in(1))) {
1950 _scale *= -1;
1951 return true;
1952 }
1953 }
1954 return false;
1955 }
1956
1957 //----------------------------scaled_iv------------------------
1958 // Match: k*iv where k is a constant that's not zero
1959 bool SWPointer::scaled_iv(Node* n) {
1960 if (_scale != 0) {
1961 return false; // already found a scale
1962 }
1963 if (n == iv()) {
1964 _scale = 1;
1965 return true;
1966 }
1967 int opc = n->Opcode();
1968 if (opc == Op_MulI) {
1969 if (n->in(1) == iv() && n->in(2)->is_Con()) {
1970 _scale = n->in(2)->get_int();
1971 return true;
1972 } else if (n->in(2) == iv() && n->in(1)->is_Con()) {
1973 _scale = n->in(1)->get_int();
1974 return true;
1975 }
1976 } else if (opc == Op_LShiftI) {
1977 if (n->in(1) == iv() && n->in(2)->is_Con()) {
1978 _scale = 1 << n->in(2)->get_int();
1979 return true;
1980 }
1981 } else if (opc == Op_ConvI2L) {
1982 if (scaled_iv_plus_offset(n->in(1))) {
1983 return true;
1984 }
1985 } else if (opc == Op_LShiftL) {
1986 if (!has_iv() && _invar == NULL) {
1987 // Need to preserve the current _offset value, so
1988 // create a temporary object for this expression subtree.
1989 // Hacky, so should re-engineer the address pattern match.
1990 SWPointer tmp(this);
1991 if (tmp.scaled_iv_plus_offset(n->in(1))) {
1992 if (tmp._invar == NULL) {
1993 int mult = 1 << n->in(2)->get_int();
1994 _scale = tmp._scale * mult;
1995 _offset += tmp._offset * mult;
1996 return true;
1997 }
1998 }
1999 }
2000 }
2001 return false;
2002 }
2003
2004 //----------------------------offset_plus_k------------------------
2005 // Match: offset is (k [+/- invariant])
2006 // where k maybe zero and invariant is optional, but not both.
2007 bool SWPointer::offset_plus_k(Node* n, bool negate) {
2008 int opc = n->Opcode();
2009 if (opc == Op_ConI) {
2010 _offset += negate ? -(n->get_int()) : n->get_int();
2011 return true;
2012 } else if (opc == Op_ConL) {
2013 // Okay if value fits into an int
2014 const TypeLong* t = n->find_long_type();
2015 if (t->higher_equal(TypeLong::INT)) {
2016 jlong loff = n->get_long();
2017 jint off = (jint)loff;
2018 _offset += negate ? -off : loff;
2019 return true;
2020 }
2021 return false;
2022 }
2023 if (_invar != NULL) return false; // already have an invariant
2024 if (opc == Op_AddI) {
2025 if (n->in(2)->is_Con() && invariant(n->in(1))) {
2026 _negate_invar = negate;
2027 _invar = n->in(1);
2028 _offset += negate ? -(n->in(2)->get_int()) : n->in(2)->get_int();
2029 return true;
2030 } else if (n->in(1)->is_Con() && invariant(n->in(2))) {
2031 _offset += negate ? -(n->in(1)->get_int()) : n->in(1)->get_int();
2032 _negate_invar = negate;
2033 _invar = n->in(2);
2034 return true;
2035 }
2036 }
2037 if (opc == Op_SubI) {
2038 if (n->in(2)->is_Con() && invariant(n->in(1))) {
2039 _negate_invar = negate;
2040 _invar = n->in(1);
2041 _offset += !negate ? -(n->in(2)->get_int()) : n->in(2)->get_int();
2042 return true;
2043 } else if (n->in(1)->is_Con() && invariant(n->in(2))) {
2044 _offset += negate ? -(n->in(1)->get_int()) : n->in(1)->get_int();
2045 _negate_invar = !negate;
2046 _invar = n->in(2);
2047 return true;
2048 }
2049 }
2050 if (invariant(n)) {
2051 _negate_invar = negate;
2052 _invar = n;
2053 return true;
2054 }
2055 return false;
2056 }
2057
2058 //----------------------------print------------------------
2059 void SWPointer::print() {
2060 #ifndef PRODUCT
2061 tty->print("base: %d adr: %d scale: %d offset: %d invar: %c%d\n",
2062 _base != NULL ? _base->_idx : 0,
2063 _adr != NULL ? _adr->_idx : 0,
2064 _scale, _offset,
2065 _negate_invar?'-':'+',
2066 _invar != NULL ? _invar->_idx : 0);
2067 #endif
2068 }
2069
2070 // ========================= OrderedPair =====================
2071
2072 const OrderedPair OrderedPair::initial;
2073
2074 // ========================= SWNodeInfo =====================
2075
2076 const SWNodeInfo SWNodeInfo::initial;
2077
2078
2079 // ============================ DepGraph ===========================
2080
2081 //------------------------------make_node---------------------------
2082 // Make a new dependence graph node for an ideal node.
2083 DepMem* DepGraph::make_node(Node* node) {
2084 DepMem* m = new (_arena) DepMem(node);
2085 if (node != NULL) {
2086 assert(_map.at_grow(node->_idx) == NULL, "one init only");
2087 _map.at_put_grow(node->_idx, m);
2088 }
2089 return m;
2090 }
2091
2092 //------------------------------make_edge---------------------------
2093 // Make a new dependence graph edge from dpred -> dsucc
2094 DepEdge* DepGraph::make_edge(DepMem* dpred, DepMem* dsucc) {
2095 DepEdge* e = new (_arena) DepEdge(dpred, dsucc, dsucc->in_head(), dpred->out_head());
2096 dpred->set_out_head(e);
2097 dsucc->set_in_head(e);
2098 return e;
2099 }
2100
2101 // ========================== DepMem ========================
2102
2103 //------------------------------in_cnt---------------------------
2104 int DepMem::in_cnt() {
2105 int ct = 0;
2106 for (DepEdge* e = _in_head; e != NULL; e = e->next_in()) ct++;
2107 return ct;
2108 }
2109
2110 //------------------------------out_cnt---------------------------
2111 int DepMem::out_cnt() {
2112 int ct = 0;
2113 for (DepEdge* e = _out_head; e != NULL; e = e->next_out()) ct++;
2114 return ct;
2115 }
2116
2117 //------------------------------print-----------------------------
2118 void DepMem::print() {
2119 #ifndef PRODUCT
2120 tty->print(" DepNode %d (", _node->_idx);
2121 for (DepEdge* p = _in_head; p != NULL; p = p->next_in()) {
2122 Node* pred = p->pred()->node();
2123 tty->print(" %d", pred != NULL ? pred->_idx : 0);
2124 }
2125 tty->print(") [");
2126 for (DepEdge* s = _out_head; s != NULL; s = s->next_out()) {
2127 Node* succ = s->succ()->node();
2128 tty->print(" %d", succ != NULL ? succ->_idx : 0);
2129 }
2130 tty->print_cr(" ]");
2131 #endif
2132 }
2133
2134 // =========================== DepEdge =========================
2135
2136 //------------------------------DepPreds---------------------------
2137 void DepEdge::print() {
2138 #ifndef PRODUCT
2139 tty->print_cr("DepEdge: %d [ %d ]", _pred->node()->_idx, _succ->node()->_idx);
2140 #endif
2141 }
2142
2143 // =========================== DepPreds =========================
2144 // Iterator over predecessor edges in the dependence graph.
2145
2146 //------------------------------DepPreds---------------------------
2147 DepPreds::DepPreds(Node* n, DepGraph& dg) {
2148 _n = n;
2149 _done = false;
2150 if (_n->is_Store() || _n->is_Load()) {
2151 _next_idx = MemNode::Address;
2152 _end_idx = n->req();
2153 _dep_next = dg.dep(_n)->in_head();
2154 } else if (_n->is_Mem()) {
2155 _next_idx = 0;
2156 _end_idx = 0;
2157 _dep_next = dg.dep(_n)->in_head();
2158 } else {
2159 _next_idx = 1;
2160 _end_idx = _n->req();
2161 _dep_next = NULL;
2162 }
2163 next();
2164 }
2165
2166 //------------------------------next---------------------------
2167 void DepPreds::next() {
2168 if (_dep_next != NULL) {
2169 _current = _dep_next->pred()->node();
2170 _dep_next = _dep_next->next_in();
2171 } else if (_next_idx < _end_idx) {
2172 _current = _n->in(_next_idx++);
2173 } else {
2174 _done = true;
2175 }
2176 }
2177
2178 // =========================== DepSuccs =========================
2179 // Iterator over successor edges in the dependence graph.
2180
2181 //------------------------------DepSuccs---------------------------
2182 DepSuccs::DepSuccs(Node* n, DepGraph& dg) {
2183 _n = n;
2184 _done = false;
2185 if (_n->is_Load()) {
2186 _next_idx = 0;
2187 _end_idx = _n->outcnt();
2188 _dep_next = dg.dep(_n)->out_head();
2189 } else if (_n->is_Mem() || _n->is_Phi() && _n->bottom_type() == Type::MEMORY) {
2190 _next_idx = 0;
2191 _end_idx = 0;
2192 _dep_next = dg.dep(_n)->out_head();
2193 } else {
2194 _next_idx = 0;
2195 _end_idx = _n->outcnt();
2196 _dep_next = NULL;
2197 }
2198 next();
2199 }
2200
2201 //-------------------------------next---------------------------
2202 void DepSuccs::next() {
2203 if (_dep_next != NULL) {
2204 _current = _dep_next->succ()->node();
2205 _dep_next = _dep_next->next_out();
2206 } else if (_next_idx < _end_idx) {
2207 _current = _n->raw_out(_next_idx++);
2208 } else {
2209 _done = true;
2210 }
2211 }