1 /*
2 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 */
23
24 #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 {
461 assert(out == prev || prev == NULL, "no branches off of store slice");
462 }
463 }
464 }
465 if (n == stop) break;
466 preds.push(n);
467 prev = n;
468 n = n->in(MemNode::Memory);
469 }
470 }
471
472 //------------------------------stmts_can_pack---------------------------
473 // Can s1 and s2 be in a pack with s1 immediately preceding s2 and
474 // s1 aligned at "align"
475 bool SuperWord::stmts_can_pack(Node* s1, Node* s2, int align) {
476
477 // Do not use superword for non-primitives
478 if((s1->is_Mem() && !is_java_primitive(s1->as_Mem()->memory_type())) ||
479 (s2->is_Mem() && !is_java_primitive(s2->as_Mem()->memory_type())))
480 return false;
481
482 if (isomorphic(s1, s2)) {
483 if (independent(s1, s2)) {
484 if (!exists_at(s1, 0) && !exists_at(s2, 1)) {
485 if (!s1->is_Mem() || are_adjacent_refs(s1, s2)) {
486 int s1_align = alignment(s1);
487 int s2_align = alignment(s2);
488 if (s1_align == top_align || s1_align == align) {
489 if (s2_align == top_align || s2_align == align + data_size(s1)) {
490 return true;
491 }
492 }
493 }
494 }
495 }
496 }
497 return false;
498 }
499
500 //------------------------------exists_at---------------------------
501 // Does s exist in a pack at position pos?
502 bool SuperWord::exists_at(Node* s, uint pos) {
503 for (int i = 0; i < _packset.length(); i++) {
504 Node_List* p = _packset.at(i);
505 if (p->at(pos) == s) {
506 return true;
507 }
508 }
509 return false;
510 }
511
512 //------------------------------are_adjacent_refs---------------------------
513 // Is s1 immediately before s2 in memory?
514 bool SuperWord::are_adjacent_refs(Node* s1, Node* s2) {
515 if (!s1->is_Mem() || !s2->is_Mem()) return false;
516 if (!in_bb(s1) || !in_bb(s2)) return false;
517
518 // Do not use superword for non-primitives
519 if (!is_java_primitive(s1->as_Mem()->memory_type()) ||
520 !is_java_primitive(s2->as_Mem()->memory_type())) {
521 return false;
522 }
523
524 // FIXME - co_locate_pack fails on Stores in different mem-slices, so
525 // only pack memops that are in the same alias set until that's fixed.
526 if (_phase->C->get_alias_index(s1->as_Mem()->adr_type()) !=
527 _phase->C->get_alias_index(s2->as_Mem()->adr_type()))
528 return false;
529 SWPointer p1(s1->as_Mem(), this);
530 SWPointer p2(s2->as_Mem(), this);
531 if (p1.base() != p2.base() || !p1.comparable(p2)) return false;
532 int diff = p2.offset_in_bytes() - p1.offset_in_bytes();
533 return diff == data_size(s1);
534 }
535
536 //------------------------------isomorphic---------------------------
537 // Are s1 and s2 similar?
538 bool SuperWord::isomorphic(Node* s1, Node* s2) {
539 if (s1->Opcode() != s2->Opcode()) return false;
540 if (s1->req() != s2->req()) return false;
541 if (s1->in(0) != s2->in(0)) return false;
542 if (velt_type(s1) != velt_type(s2)) return false;
543 return true;
544 }
545
546 //------------------------------independent---------------------------
547 // Is there no data path from s1 to s2 or s2 to s1?
548 bool SuperWord::independent(Node* s1, Node* s2) {
549 // assert(s1->Opcode() == s2->Opcode(), "check isomorphic first");
550 int d1 = depth(s1);
551 int d2 = depth(s2);
552 if (d1 == d2) return s1 != s2;
553 Node* deep = d1 > d2 ? s1 : s2;
554 Node* shallow = d1 > d2 ? s2 : s1;
555
556 visited_clear();
557
558 return independent_path(shallow, deep);
559 }
560
561 //------------------------------independent_path------------------------------
562 // Helper for independent
563 bool SuperWord::independent_path(Node* shallow, Node* deep, uint dp) {
564 if (dp >= 1000) return false; // stop deep recursion
565 visited_set(deep);
566 int shal_depth = depth(shallow);
567 assert(shal_depth <= depth(deep), "must be");
568 for (DepPreds preds(deep, _dg); !preds.done(); preds.next()) {
569 Node* pred = preds.current();
570 if (in_bb(pred) && !visited_test(pred)) {
571 if (shallow == pred) {
572 return false;
573 }
574 if (shal_depth < depth(pred) && !independent_path(shallow, pred, dp+1)) {
575 return false;
576 }
577 }
578 }
579 return true;
580 }
581
582 //------------------------------set_alignment---------------------------
583 void SuperWord::set_alignment(Node* s1, Node* s2, int align) {
584 set_alignment(s1, align);
585 set_alignment(s2, align + data_size(s1));
586 }
587
588 //------------------------------data_size---------------------------
589 int SuperWord::data_size(Node* s) {
590 const Type* t = velt_type(s);
591 BasicType bt = t->array_element_basic_type();
592 int bsize = type2aelembytes(bt);
593 assert(bsize != 0, "valid size");
594 return bsize;
595 }
596
597 //------------------------------extend_packlist---------------------------
598 // Extend packset by following use->def and def->use links from pack members.
599 void SuperWord::extend_packlist() {
600 bool changed;
601 do {
602 changed = false;
603 for (int i = 0; i < _packset.length(); i++) {
604 Node_List* p = _packset.at(i);
605 changed |= follow_use_defs(p);
606 changed |= follow_def_uses(p);
607 }
608 } while (changed);
609
610 #ifndef PRODUCT
611 if (TraceSuperWord) {
612 tty->print_cr("\nAfter extend_packlist");
613 print_packset();
614 }
615 #endif
616 }
617
618 //------------------------------follow_use_defs---------------------------
619 // Extend the packset by visiting operand definitions of nodes in pack p
620 bool SuperWord::follow_use_defs(Node_List* p) {
621 Node* s1 = p->at(0);
622 Node* s2 = p->at(1);
623 assert(p->size() == 2, "just checking");
624 assert(s1->req() == s2->req(), "just checking");
625 assert(alignment(s1) + data_size(s1) == alignment(s2), "just checking");
626
627 if (s1->is_Load()) return false;
628
629 int align = alignment(s1);
630 bool changed = false;
631 int start = s1->is_Store() ? MemNode::ValueIn : 1;
632 int end = s1->is_Store() ? MemNode::ValueIn+1 : s1->req();
633 for (int j = start; j < end; j++) {
634 Node* t1 = s1->in(j);
635 Node* t2 = s2->in(j);
636 if (!in_bb(t1) || !in_bb(t2))
637 continue;
638 if (stmts_can_pack(t1, t2, align)) {
639 if (est_savings(t1, t2) >= 0) {
640 Node_List* pair = new Node_List();
641 pair->push(t1);
642 pair->push(t2);
643 _packset.append(pair);
644 set_alignment(t1, t2, align);
645 changed = true;
646 }
647 }
648 }
649 return changed;
650 }
651
652 //------------------------------follow_def_uses---------------------------
653 // Extend the packset by visiting uses of nodes in pack p
654 bool SuperWord::follow_def_uses(Node_List* p) {
655 bool changed = false;
656 Node* s1 = p->at(0);
657 Node* s2 = p->at(1);
658 assert(p->size() == 2, "just checking");
659 assert(s1->req() == s2->req(), "just checking");
660 assert(alignment(s1) + data_size(s1) == alignment(s2), "just checking");
661
662 if (s1->is_Store()) return false;
663
664 int align = alignment(s1);
665 int savings = -1;
666 Node* u1 = NULL;
667 Node* u2 = NULL;
668 for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
669 Node* t1 = s1->fast_out(i);
670 if (!in_bb(t1)) continue;
671 for (DUIterator_Fast jmax, j = s2->fast_outs(jmax); j < jmax; j++) {
672 Node* t2 = s2->fast_out(j);
673 if (!in_bb(t2)) continue;
674 if (!opnd_positions_match(s1, t1, s2, t2))
675 continue;
676 if (stmts_can_pack(t1, t2, align)) {
677 int my_savings = est_savings(t1, t2);
678 if (my_savings > savings) {
679 savings = my_savings;
680 u1 = t1;
681 u2 = t2;
682 }
683 }
684 }
685 }
686 if (savings >= 0) {
687 Node_List* pair = new Node_List();
688 pair->push(u1);
689 pair->push(u2);
690 _packset.append(pair);
691 set_alignment(u1, u2, align);
692 changed = true;
693 }
694 return changed;
695 }
696
697 //---------------------------opnd_positions_match-------------------------
698 // Is the use of d1 in u1 at the same operand position as d2 in u2?
699 bool SuperWord::opnd_positions_match(Node* d1, Node* u1, Node* d2, Node* u2) {
700 uint ct = u1->req();
701 if (ct != u2->req()) return false;
702 uint i1 = 0;
703 uint i2 = 0;
704 do {
705 for (i1++; i1 < ct; i1++) if (u1->in(i1) == d1) break;
706 for (i2++; i2 < ct; i2++) if (u2->in(i2) == d2) break;
707 if (i1 != i2) {
708 return false;
709 }
710 } while (i1 < ct);
711 return true;
712 }
713
714 //------------------------------est_savings---------------------------
715 // Estimate the savings from executing s1 and s2 as a pack
716 int SuperWord::est_savings(Node* s1, Node* s2) {
717 int save = 2 - 1; // 2 operations per instruction in packed form
718
719 // inputs
720 for (uint i = 1; i < s1->req(); i++) {
721 Node* x1 = s1->in(i);
722 Node* x2 = s2->in(i);
723 if (x1 != x2) {
724 if (are_adjacent_refs(x1, x2)) {
725 save += adjacent_profit(x1, x2);
726 } else if (!in_packset(x1, x2)) {
727 save -= pack_cost(2);
728 } else {
729 save += unpack_cost(2);
730 }
731 }
732 }
733
734 // uses of result
735 uint ct = 0;
736 for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
737 Node* s1_use = s1->fast_out(i);
738 for (int j = 0; j < _packset.length(); j++) {
739 Node_List* p = _packset.at(j);
740 if (p->at(0) == s1_use) {
741 for (DUIterator_Fast kmax, k = s2->fast_outs(kmax); k < kmax; k++) {
742 Node* s2_use = s2->fast_out(k);
743 if (p->at(p->size()-1) == s2_use) {
744 ct++;
745 if (are_adjacent_refs(s1_use, s2_use)) {
746 save += adjacent_profit(s1_use, s2_use);
747 }
748 }
749 }
750 }
751 }
752 }
753
754 if (ct < s1->outcnt()) save += unpack_cost(1);
755 if (ct < s2->outcnt()) save += unpack_cost(1);
756
757 return save;
758 }
759
760 //------------------------------costs---------------------------
761 int SuperWord::adjacent_profit(Node* s1, Node* s2) { return 2; }
762 int SuperWord::pack_cost(int ct) { return ct; }
763 int SuperWord::unpack_cost(int ct) { return ct; }
764
765 //------------------------------combine_packs---------------------------
766 // Combine packs A and B with A.last == B.first into A.first..,A.last,B.second,..B.last
767 void SuperWord::combine_packs() {
768 bool changed;
769 do {
770 changed = false;
771 for (int i = 0; i < _packset.length(); i++) {
772 Node_List* p1 = _packset.at(i);
773 if (p1 == NULL) continue;
774 for (int j = 0; j < _packset.length(); j++) {
775 Node_List* p2 = _packset.at(j);
776 if (p2 == NULL) continue;
777 if (p1->at(p1->size()-1) == p2->at(0)) {
778 for (uint k = 1; k < p2->size(); k++) {
779 p1->push(p2->at(k));
780 }
781 _packset.at_put(j, NULL);
782 changed = true;
783 }
784 }
785 }
786 } while (changed);
787
788 for (int i = _packset.length() - 1; i >= 0; i--) {
789 Node_List* p1 = _packset.at(i);
790 if (p1 == NULL) {
791 _packset.remove_at(i);
792 }
793 }
794
795 #ifndef PRODUCT
796 if (TraceSuperWord) {
797 tty->print_cr("\nAfter combine_packs");
798 print_packset();
799 }
800 #endif
801 }
802
803 //-----------------------------construct_my_pack_map--------------------------
804 // Construct the map from nodes to packs. Only valid after the
805 // point where a node is only in one pack (after combine_packs).
806 void SuperWord::construct_my_pack_map() {
807 Node_List* rslt = NULL;
808 for (int i = 0; i < _packset.length(); i++) {
809 Node_List* p = _packset.at(i);
810 for (uint j = 0; j < p->size(); j++) {
811 Node* s = p->at(j);
812 assert(my_pack(s) == NULL, "only in one pack");
813 set_my_pack(s, p);
814 }
815 }
816 }
817
818 //------------------------------filter_packs---------------------------
819 // Remove packs that are not implemented or not profitable.
820 void SuperWord::filter_packs() {
821
822 // Remove packs that are not implemented
823 for (int i = _packset.length() - 1; i >= 0; i--) {
824 Node_List* pk = _packset.at(i);
825 bool impl = implemented(pk);
826 if (!impl) {
827 #ifndef PRODUCT
828 if (TraceSuperWord && Verbose) {
829 tty->print_cr("Unimplemented");
830 pk->at(0)->dump();
831 }
832 #endif
833 remove_pack_at(i);
834 }
835 }
836
837 // Remove packs that are not profitable
838 bool changed;
839 do {
840 changed = false;
841 for (int i = _packset.length() - 1; i >= 0; i--) {
842 Node_List* pk = _packset.at(i);
843 bool prof = profitable(pk);
844 if (!prof) {
845 #ifndef PRODUCT
846 if (TraceSuperWord && Verbose) {
847 tty->print_cr("Unprofitable");
848 pk->at(0)->dump();
849 }
850 #endif
851 remove_pack_at(i);
852 changed = true;
853 }
854 }
855 } while (changed);
856
857 #ifndef PRODUCT
858 if (TraceSuperWord) {
859 tty->print_cr("\nAfter filter_packs");
860 print_packset();
861 tty->cr();
862 }
863 #endif
864 }
865
866 //------------------------------implemented---------------------------
867 // Can code be generated for pack p?
868 bool SuperWord::implemented(Node_List* p) {
869 Node* p0 = p->at(0);
870 int vopc = VectorNode::opcode(p0->Opcode(), p->size(), velt_type(p0));
871 return vopc > 0 && Matcher::has_match_rule(vopc);
872 }
873
874 //------------------------------profitable---------------------------
875 // For pack p, are all operands and all uses (with in the block) vector?
876 bool SuperWord::profitable(Node_List* p) {
877 Node* p0 = p->at(0);
878 uint start, end;
879 vector_opd_range(p0, &start, &end);
880
881 // Return false if some input is not vector and inside block
882 for (uint i = start; i < end; i++) {
883 if (!is_vector_use(p0, i)) {
884 // For now, return false if not scalar promotion case (inputs are the same.)
885 // Later, implement PackNode and allow differing, non-vector inputs
886 // (maybe just the ones from outside the block.)
887 Node* p0_def = p0->in(i);
888 for (uint j = 1; j < p->size(); j++) {
889 Node* use = p->at(j);
890 Node* def = use->in(i);
891 if (p0_def != def)
892 return false;
893 }
894 }
895 }
896 if (!p0->is_Store()) {
897 // For now, return false if not all uses are vector.
898 // Later, implement ExtractNode and allow non-vector uses (maybe
899 // just the ones outside the block.)
900 for (uint i = 0; i < p->size(); i++) {
901 Node* def = p->at(i);
902 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
903 Node* use = def->fast_out(j);
904 for (uint k = 0; k < use->req(); k++) {
905 Node* n = use->in(k);
906 if (def == n) {
907 if (!is_vector_use(use, k)) {
908 return false;
909 }
910 }
911 }
912 }
913 }
914 }
915 return true;
916 }
917
918 //------------------------------schedule---------------------------
919 // Adjust the memory graph for the packed operations
920 void SuperWord::schedule() {
921
922 // Co-locate in the memory graph the members of each memory pack
923 for (int i = 0; i < _packset.length(); i++) {
924 co_locate_pack(_packset.at(i));
925 }
926 }
927
928 //-------------------------------remove_and_insert-------------------
929 //remove "current" from its current position in the memory graph and insert
930 //it after the appropriate insertion point (lip or uip)
931 void SuperWord::remove_and_insert(MemNode *current, MemNode *prev, MemNode *lip,
932 Node *uip, Unique_Node_List &sched_before) {
933 Node* my_mem = current->in(MemNode::Memory);
934 _igvn.hash_delete(current);
935 _igvn.hash_delete(my_mem);
936
937 //remove current_store from its current position in the memmory graph
938 for (DUIterator i = current->outs(); current->has_out(i); i++) {
939 Node* use = current->out(i);
940 if (use->is_Mem()) {
941 assert(use->in(MemNode::Memory) == current, "must be");
942 _igvn.hash_delete(use);
943 if (use == prev) { // connect prev to my_mem
944 use->set_req(MemNode::Memory, my_mem);
945 } else if (sched_before.member(use)) {
946 _igvn.hash_delete(uip);
947 use->set_req(MemNode::Memory, uip);
948 } else {
949 _igvn.hash_delete(lip);
950 use->set_req(MemNode::Memory, lip);
951 }
952 _igvn._worklist.push(use);
953 --i; //deleted this edge; rescan position
954 }
955 }
956
957 bool sched_up = sched_before.member(current);
958 Node *insert_pt = sched_up ? uip : lip;
959 _igvn.hash_delete(insert_pt);
960
961 // all uses of insert_pt's memory state should use current's instead
962 for (DUIterator i = insert_pt->outs(); insert_pt->has_out(i); i++) {
963 Node* use = insert_pt->out(i);
964 if (use->is_Mem()) {
965 assert(use->in(MemNode::Memory) == insert_pt, "must be");
966 _igvn.hash_delete(use);
967 use->set_req(MemNode::Memory, current);
968 _igvn._worklist.push(use);
969 --i; //deleted this edge; rescan position
970 } else if (!sched_up && use->is_Phi() && use->bottom_type() == Type::MEMORY) {
971 uint pos; //lip (lower insert point) must be the last one in the memory slice
972 _igvn.hash_delete(use);
973 for (pos=1; pos < use->req(); pos++) {
974 if (use->in(pos) == insert_pt) break;
975 }
976 use->set_req(pos, current);
977 _igvn._worklist.push(use);
978 --i;
979 }
980 }
981
982 //connect current to insert_pt
983 current->set_req(MemNode::Memory, insert_pt);
984 _igvn._worklist.push(current);
985 }
986
987 //------------------------------co_locate_pack----------------------------------
988 // To schedule a store pack, we need to move any sandwiched memory ops either before
989 // or after the pack, based upon dependence information:
990 // (1) If any store in the pack depends on the sandwiched memory op, the
991 // sandwiched memory op must be scheduled BEFORE the pack;
992 // (2) If a sandwiched memory op depends on any store in the pack, the
993 // sandwiched memory op must be scheduled AFTER the pack;
994 // (3) If a sandwiched memory op (say, memA) depends on another sandwiched
995 // memory op (say memB), memB must be scheduled before memA. So, if memA is
996 // scheduled before the pack, memB must also be scheduled before the pack;
997 // (4) If there is no dependence restriction for a sandwiched memory op, we simply
998 // schedule this store AFTER the pack
999 // (5) We know there is no dependence cycle, so there in no other case;
1000 // (6) Finally, all memory ops in another single pack should be moved in the same direction.
1001 //
1002 // To schedule a load pack, we use the memory state of either the first or the last load in
1003 // the pack, based on the dependence constraint.
1004 void SuperWord::co_locate_pack(Node_List* pk) {
1005 if (pk->at(0)->is_Store()) {
1006 MemNode* first = executed_first(pk)->as_Mem();
1007 MemNode* last = executed_last(pk)->as_Mem();
1008 Unique_Node_List schedule_before_pack;
1009 Unique_Node_List memops;
1010
1011 MemNode* current = last->in(MemNode::Memory)->as_Mem();
1012 MemNode* previous = last;
1013 while (true) {
1014 assert(in_bb(current), "stay in block");
1015 memops.push(previous);
1016 for (DUIterator i = current->outs(); current->has_out(i); i++) {
1017 Node* use = current->out(i);
1018 if (use->is_Mem() && use != previous)
1019 memops.push(use);
1020 }
1021 if(current == first) break;
1022 previous = current;
1023 current = current->in(MemNode::Memory)->as_Mem();
1024 }
1025
1026 // determine which memory operations should be scheduled before the pack
1027 for (uint i = 1; i < memops.size(); i++) {
1028 Node *s1 = memops.at(i);
1029 if (!in_pack(s1, pk) && !schedule_before_pack.member(s1)) {
1030 for (uint j = 0; j< i; j++) {
1031 Node *s2 = memops.at(j);
1032 if (!independent(s1, s2)) {
1033 if (in_pack(s2, pk) || schedule_before_pack.member(s2)) {
1034 schedule_before_pack.push(s1); //s1 must be scheduled before
1035 Node_List* mem_pk = my_pack(s1);
1036 if (mem_pk != NULL) {
1037 for (uint ii = 0; ii < mem_pk->size(); ii++) {
1038 Node* s = mem_pk->at(ii); // follow partner
1039 if (memops.member(s) && !schedule_before_pack.member(s))
1040 schedule_before_pack.push(s);
1041 }
1042 }
1043 }
1044 }
1045 }
1046 }
1047 }
1048
1049 MemNode* lower_insert_pt = last;
1050 Node* upper_insert_pt = first->in(MemNode::Memory);
1051 previous = last; //previous store in pk
1052 current = last->in(MemNode::Memory)->as_Mem();
1053
1054 //start scheduling from "last" to "first"
1055 while (true) {
1056 assert(in_bb(current), "stay in block");
1057 assert(in_pack(previous, pk), "previous stays in pack");
1058 Node* my_mem = current->in(MemNode::Memory);
1059
1060 if (in_pack(current, pk)) {
1061 // Forward users of my memory state (except "previous) to my input memory state
1062 _igvn.hash_delete(current);
1063 for (DUIterator i = current->outs(); current->has_out(i); i++) {
1064 Node* use = current->out(i);
1065 if (use->is_Mem() && use != previous) {
1066 assert(use->in(MemNode::Memory) == current, "must be");
1067 _igvn.hash_delete(use);
1068 if (schedule_before_pack.member(use)) {
1069 _igvn.hash_delete(upper_insert_pt);
1070 use->set_req(MemNode::Memory, upper_insert_pt);
1071 } else {
1072 _igvn.hash_delete(lower_insert_pt);
1073 use->set_req(MemNode::Memory, lower_insert_pt);
1074 }
1075 _igvn._worklist.push(use);
1076 --i; // deleted this edge; rescan position
1077 }
1078 }
1079 previous = current;
1080 } else { // !in_pack(current, pk) ==> a sandwiched store
1081 remove_and_insert(current, previous, lower_insert_pt, upper_insert_pt, schedule_before_pack);
1082 }
1083
1084 if (current == first) break;
1085 current = my_mem->as_Mem();
1086 } // end while
1087 } else if (pk->at(0)->is_Load()) { //load
1088 // all loads in the pack should have the same memory state. By default,
1089 // we use the memory state of the last load. However, if any load could
1090 // not be moved down due to the dependence constraint, we use the memory
1091 // state of the first load.
1092 Node* last_mem = executed_last(pk)->in(MemNode::Memory);
1093 Node* first_mem = executed_first(pk)->in(MemNode::Memory);
1094 bool schedule_last = true;
1095 for (uint i = 0; i < pk->size(); i++) {
1096 Node* ld = pk->at(i);
1097 for (Node* current = last_mem; current != ld->in(MemNode::Memory);
1098 current=current->in(MemNode::Memory)) {
1099 assert(current != first_mem, "corrupted memory graph");
1100 if(current->is_Mem() && !independent(current, ld)){
1101 schedule_last = false; // a later store depends on this load
1102 break;
1103 }
1104 }
1105 }
1106
1107 Node* mem_input = schedule_last ? last_mem : first_mem;
1108 _igvn.hash_delete(mem_input);
1109 // Give each load the same memory state
1110 for (uint i = 0; i < pk->size(); i++) {
1111 LoadNode* ld = pk->at(i)->as_Load();
1112 _igvn.hash_delete(ld);
1113 ld->set_req(MemNode::Memory, mem_input);
1114 _igvn._worklist.push(ld);
1115 }
1116 }
1117 }
1118
1119 //------------------------------output---------------------------
1120 // Convert packs into vector node operations
1121 void SuperWord::output() {
1122 if (_packset.length() == 0) return;
1123
1124 // MUST ENSURE main loop's initial value is properly aligned:
1125 // (iv_initial_value + min_iv_offset) % vector_width_in_bytes() == 0
1126
1127 align_initial_loop_index(align_to_ref());
1128
1129 // Insert extract (unpack) operations for scalar uses
1130 for (int i = 0; i < _packset.length(); i++) {
1131 insert_extracts(_packset.at(i));
1132 }
1133
1134 for (int i = 0; i < _block.length(); i++) {
1135 Node* n = _block.at(i);
1136 Node_List* p = my_pack(n);
1137 if (p && n == executed_last(p)) {
1138 uint vlen = p->size();
1139 Node* vn = NULL;
1140 Node* low_adr = p->at(0);
1141 Node* first = executed_first(p);
1142 if (n->is_Load()) {
1143 int opc = n->Opcode();
1144 Node* ctl = n->in(MemNode::Control);
1145 Node* mem = first->in(MemNode::Memory);
1146 Node* adr = low_adr->in(MemNode::Address);
1147 const TypePtr* atyp = n->adr_type();
1148 vn = VectorLoadNode::make(_phase->C, opc, ctl, mem, adr, atyp, vlen);
1149
1150 } else if (n->is_Store()) {
1151 // Promote value to be stored to vector
1152 VectorNode* val = vector_opd(p, MemNode::ValueIn);
1153
1154 int opc = n->Opcode();
1155 Node* ctl = n->in(MemNode::Control);
1156 Node* mem = first->in(MemNode::Memory);
1157 Node* adr = low_adr->in(MemNode::Address);
1158 const TypePtr* atyp = n->adr_type();
1159 vn = VectorStoreNode::make(_phase->C, opc, ctl, mem, adr, atyp, val, vlen);
1160
1161 } else if (n->req() == 3) {
1162 // Promote operands to vector
1163 Node* in1 = vector_opd(p, 1);
1164 Node* in2 = vector_opd(p, 2);
1165 vn = VectorNode::make(_phase->C, n->Opcode(), in1, in2, vlen, velt_type(n));
1166
1167 } else {
1168 ShouldNotReachHere();
1169 }
1170
1171 _phase->_igvn.register_new_node_with_optimizer(vn);
1172 _phase->set_ctrl(vn, _phase->get_ctrl(p->at(0)));
1173 for (uint j = 0; j < p->size(); j++) {
1174 Node* pm = p->at(j);
1175 _igvn.replace_node(pm, vn);
1176 }
1177 _igvn._worklist.push(vn);
1178 }
1179 }
1180 }
1181
1182 //------------------------------vector_opd---------------------------
1183 // Create a vector operand for the nodes in pack p for operand: in(opd_idx)
1184 VectorNode* SuperWord::vector_opd(Node_List* p, int opd_idx) {
1185 Node* p0 = p->at(0);
1186 uint vlen = p->size();
1187 Node* opd = p0->in(opd_idx);
1188
1189 bool same_opd = true;
1190 for (uint i = 1; i < vlen; i++) {
1191 Node* pi = p->at(i);
1192 Node* in = pi->in(opd_idx);
1193 if (opd != in) {
1194 same_opd = false;
1195 break;
1196 }
1197 }
1198
1199 if (same_opd) {
1200 if (opd->is_Vector()) {
1201 return (VectorNode*)opd; // input is matching vector
1202 }
1203 // Convert scalar input to vector. Use p0's type because it's container
1204 // maybe smaller than the operand's container.
1205 const Type* opd_t = velt_type(!in_bb(opd) ? p0 : opd);
1206 const Type* p0_t = velt_type(p0);
1207 if (p0_t->higher_equal(opd_t)) opd_t = p0_t;
1208 VectorNode* vn = VectorNode::scalar2vector(_phase->C, opd, vlen, opd_t);
1209
1210 _phase->_igvn.register_new_node_with_optimizer(vn);
1211 _phase->set_ctrl(vn, _phase->get_ctrl(opd));
1212 return vn;
1213 }
1214
1215 // Insert pack operation
1216 const Type* opd_t = velt_type(!in_bb(opd) ? p0 : opd);
1217 PackNode* pk = PackNode::make(_phase->C, opd, opd_t);
1218
1219 for (uint i = 1; i < vlen; i++) {
1220 Node* pi = p->at(i);
1221 Node* in = pi->in(opd_idx);
1222 assert(my_pack(in) == NULL, "Should already have been unpacked");
1223 assert(opd_t == velt_type(!in_bb(in) ? pi : in), "all same type");
1224 pk->add_opd(in);
1225 }
1226 _phase->_igvn.register_new_node_with_optimizer(pk);
1227 _phase->set_ctrl(pk, _phase->get_ctrl(opd));
1228 return pk;
1229 }
1230
1231 //------------------------------insert_extracts---------------------------
1232 // If a use of pack p is not a vector use, then replace the
1233 // use with an extract operation.
1234 void SuperWord::insert_extracts(Node_List* p) {
1235 if (p->at(0)->is_Store()) return;
1236 assert(_n_idx_list.is_empty(), "empty (node,index) list");
1237
1238 // Inspect each use of each pack member. For each use that is
1239 // not a vector use, replace the use with an extract operation.
1240
1241 for (uint i = 0; i < p->size(); i++) {
1242 Node* def = p->at(i);
1243 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
1244 Node* use = def->fast_out(j);
1245 for (uint k = 0; k < use->req(); k++) {
1246 Node* n = use->in(k);
1247 if (def == n) {
1248 if (!is_vector_use(use, k)) {
1249 _n_idx_list.push(use, k);
1250 }
1251 }
1252 }
1253 }
1254 }
1255
1256 while (_n_idx_list.is_nonempty()) {
1257 Node* use = _n_idx_list.node();
1258 int idx = _n_idx_list.index();
1259 _n_idx_list.pop();
1260 Node* def = use->in(idx);
1261
1262 // Insert extract operation
1263 _igvn.hash_delete(def);
1264 _igvn.hash_delete(use);
1265 int def_pos = alignment(def) / data_size(def);
1266 const Type* def_t = velt_type(def);
1267
1268 Node* ex = ExtractNode::make(_phase->C, def, def_pos, def_t);
1269 _phase->_igvn.register_new_node_with_optimizer(ex);
1270 _phase->set_ctrl(ex, _phase->get_ctrl(def));
1271 use->set_req(idx, ex);
1272 _igvn._worklist.push(def);
1273 _igvn._worklist.push(use);
1274
1275 bb_insert_after(ex, bb_idx(def));
1276 set_velt_type(ex, def_t);
1277 }
1278 }
1279
1280 //------------------------------is_vector_use---------------------------
1281 // Is use->in(u_idx) a vector use?
1282 bool SuperWord::is_vector_use(Node* use, int u_idx) {
1283 Node_List* u_pk = my_pack(use);
1284 if (u_pk == NULL) return false;
1285 Node* def = use->in(u_idx);
1286 Node_List* d_pk = my_pack(def);
1287 if (d_pk == NULL) {
1288 // check for scalar promotion
1289 Node* n = u_pk->at(0)->in(u_idx);
1290 for (uint i = 1; i < u_pk->size(); i++) {
1291 if (u_pk->at(i)->in(u_idx) != n) return false;
1292 }
1293 return true;
1294 }
1295 if (u_pk->size() != d_pk->size())
1296 return false;
1297 for (uint i = 0; i < u_pk->size(); i++) {
1298 Node* ui = u_pk->at(i);
1299 Node* di = d_pk->at(i);
1300 if (ui->in(u_idx) != di || alignment(ui) != alignment(di))
1301 return false;
1302 }
1303 return true;
1304 }
1305
1306 //------------------------------construct_bb---------------------------
1307 // Construct reverse postorder list of block members
1308 void SuperWord::construct_bb() {
1309 Node* entry = bb();
1310
1311 assert(_stk.length() == 0, "stk is empty");
1312 assert(_block.length() == 0, "block is empty");
1313 assert(_data_entry.length() == 0, "data_entry is empty");
1314 assert(_mem_slice_head.length() == 0, "mem_slice_head is empty");
1315 assert(_mem_slice_tail.length() == 0, "mem_slice_tail is empty");
1316
1317 // Find non-control nodes with no inputs from within block,
1318 // create a temporary map from node _idx to bb_idx for use
1319 // by the visited and post_visited sets,
1320 // and count number of nodes in block.
1321 int bb_ct = 0;
1322 for (uint i = 0; i < lpt()->_body.size(); i++ ) {
1323 Node *n = lpt()->_body.at(i);
1324 set_bb_idx(n, i); // Create a temporary map
1325 if (in_bb(n)) {
1326 bb_ct++;
1327 if (!n->is_CFG()) {
1328 bool found = false;
1329 for (uint j = 0; j < n->req(); j++) {
1330 Node* def = n->in(j);
1331 if (def && in_bb(def)) {
1332 found = true;
1333 break;
1334 }
1335 }
1336 if (!found) {
1337 assert(n != entry, "can't be entry");
1338 _data_entry.push(n);
1339 }
1340 }
1341 }
1342 }
1343
1344 // Find memory slices (head and tail)
1345 for (DUIterator_Fast imax, i = lp()->fast_outs(imax); i < imax; i++) {
1346 Node *n = lp()->fast_out(i);
1347 if (in_bb(n) && (n->is_Phi() && n->bottom_type() == Type::MEMORY)) {
1348 Node* n_tail = n->in(LoopNode::LoopBackControl);
1349 if (n_tail != n->in(LoopNode::EntryControl)) {
1350 _mem_slice_head.push(n);
1351 _mem_slice_tail.push(n_tail);
1352 }
1353 }
1354 }
1355
1356 // Create an RPO list of nodes in block
1357
1358 visited_clear();
1359 post_visited_clear();
1360
1361 // Push all non-control nodes with no inputs from within block, then control entry
1362 for (int j = 0; j < _data_entry.length(); j++) {
1363 Node* n = _data_entry.at(j);
1364 visited_set(n);
1365 _stk.push(n);
1366 }
1367 visited_set(entry);
1368 _stk.push(entry);
1369
1370 // Do a depth first walk over out edges
1371 int rpo_idx = bb_ct - 1;
1372 int size;
1373 while ((size = _stk.length()) > 0) {
1374 Node* n = _stk.top(); // Leave node on stack
1375 if (!visited_test_set(n)) {
1376 // forward arc in graph
1377 } else if (!post_visited_test(n)) {
1378 // cross or back arc
1379 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1380 Node *use = n->fast_out(i);
1381 if (in_bb(use) && !visited_test(use) &&
1382 // Don't go around backedge
1383 (!use->is_Phi() || n == entry)) {
1384 _stk.push(use);
1385 }
1386 }
1387 if (_stk.length() == size) {
1388 // There were no additional uses, post visit node now
1389 _stk.pop(); // Remove node from stack
1390 assert(rpo_idx >= 0, "");
1391 _block.at_put_grow(rpo_idx, n);
1392 rpo_idx--;
1393 post_visited_set(n);
1394 assert(rpo_idx >= 0 || _stk.is_empty(), "");
1395 }
1396 } else {
1397 _stk.pop(); // Remove post-visited node from stack
1398 }
1399 }
1400
1401 // Create real map of block indices for nodes
1402 for (int j = 0; j < _block.length(); j++) {
1403 Node* n = _block.at(j);
1404 set_bb_idx(n, j);
1405 }
1406
1407 initialize_bb(); // Ensure extra info is allocated.
1408
1409 #ifndef PRODUCT
1410 if (TraceSuperWord) {
1411 print_bb();
1412 tty->print_cr("\ndata entry nodes: %s", _data_entry.length() > 0 ? "" : "NONE");
1413 for (int m = 0; m < _data_entry.length(); m++) {
1414 tty->print("%3d ", m);
1415 _data_entry.at(m)->dump();
1416 }
1417 tty->print_cr("\nmemory slices: %s", _mem_slice_head.length() > 0 ? "" : "NONE");
1418 for (int m = 0; m < _mem_slice_head.length(); m++) {
1419 tty->print("%3d ", m); _mem_slice_head.at(m)->dump();
1420 tty->print(" "); _mem_slice_tail.at(m)->dump();
1421 }
1422 }
1423 #endif
1424 assert(rpo_idx == -1 && bb_ct == _block.length(), "all block members found");
1425 }
1426
1427 //------------------------------initialize_bb---------------------------
1428 // Initialize per node info
1429 void SuperWord::initialize_bb() {
1430 Node* last = _block.at(_block.length() - 1);
1431 grow_node_info(bb_idx(last));
1432 }
1433
1434 //------------------------------bb_insert_after---------------------------
1435 // Insert n into block after pos
1436 void SuperWord::bb_insert_after(Node* n, int pos) {
1437 int n_pos = pos + 1;
1438 // Make room
1439 for (int i = _block.length() - 1; i >= n_pos; i--) {
1440 _block.at_put_grow(i+1, _block.at(i));
1441 }
1442 for (int j = _node_info.length() - 1; j >= n_pos; j--) {
1443 _node_info.at_put_grow(j+1, _node_info.at(j));
1444 }
1445 // Set value
1446 _block.at_put_grow(n_pos, n);
1447 _node_info.at_put_grow(n_pos, SWNodeInfo::initial);
1448 // Adjust map from node->_idx to _block index
1449 for (int i = n_pos; i < _block.length(); i++) {
1450 set_bb_idx(_block.at(i), i);
1451 }
1452 }
1453
1454 //------------------------------compute_max_depth---------------------------
1455 // Compute max depth for expressions from beginning of block
1456 // Use to prune search paths during test for independence.
1457 void SuperWord::compute_max_depth() {
1458 int ct = 0;
1459 bool again;
1460 do {
1461 again = false;
1462 for (int i = 0; i < _block.length(); i++) {
1463 Node* n = _block.at(i);
1464 if (!n->is_Phi()) {
1465 int d_orig = depth(n);
1466 int d_in = 0;
1467 for (DepPreds preds(n, _dg); !preds.done(); preds.next()) {
1468 Node* pred = preds.current();
1469 if (in_bb(pred)) {
1470 d_in = MAX2(d_in, depth(pred));
1471 }
1472 }
1473 if (d_in + 1 != d_orig) {
1474 set_depth(n, d_in + 1);
1475 again = true;
1476 }
1477 }
1478 }
1479 ct++;
1480 } while (again);
1481 #ifndef PRODUCT
1482 if (TraceSuperWord && Verbose)
1483 tty->print_cr("compute_max_depth iterated: %d times", ct);
1484 #endif
1485 }
1486
1487 //-------------------------compute_vector_element_type-----------------------
1488 // Compute necessary vector element type for expressions
1489 // This propagates backwards a narrower integer type when the
1490 // upper bits of the value are not needed.
1491 // Example: char a,b,c; a = b + c;
1492 // Normally the type of the add is integer, but for packed character
1493 // operations the type of the add needs to be char.
1494 void SuperWord::compute_vector_element_type() {
1495 #ifndef PRODUCT
1496 if (TraceSuperWord && Verbose)
1497 tty->print_cr("\ncompute_velt_type:");
1498 #endif
1499
1500 // Initial type
1501 for (int i = 0; i < _block.length(); i++) {
1502 Node* n = _block.at(i);
1503 const Type* t = n->is_Mem() ? Type::get_const_basic_type(n->as_Mem()->memory_type())
1504 : _igvn.type(n);
1505 const Type* vt = container_type(t);
1506 set_velt_type(n, vt);
1507 }
1508
1509 // Propagate narrowed type backwards through operations
1510 // that don't depend on higher order bits
1511 for (int i = _block.length() - 1; i >= 0; i--) {
1512 Node* n = _block.at(i);
1513 // Only integer types need be examined
1514 if (n->bottom_type()->isa_int()) {
1515 uint start, end;
1516 vector_opd_range(n, &start, &end);
1517 const Type* vt = velt_type(n);
1518
1519 for (uint j = start; j < end; j++) {
1520 Node* in = n->in(j);
1521 // Don't propagate through a type conversion
1522 if (n->bottom_type() != in->bottom_type())
1523 continue;
1524 switch(in->Opcode()) {
1525 case Op_AddI: case Op_AddL:
1526 case Op_SubI: case Op_SubL:
1527 case Op_MulI: case Op_MulL:
1528 case Op_AndI: case Op_AndL:
1529 case Op_OrI: case Op_OrL:
1530 case Op_XorI: case Op_XorL:
1531 case Op_LShiftI: case Op_LShiftL:
1532 case Op_CMoveI: case Op_CMoveL:
1533 if (in_bb(in)) {
1534 bool same_type = true;
1535 for (DUIterator_Fast kmax, k = in->fast_outs(kmax); k < kmax; k++) {
1536 Node *use = in->fast_out(k);
1537 if (!in_bb(use) || velt_type(use) != vt) {
1538 same_type = false;
1539 break;
1540 }
1541 }
1542 if (same_type) {
1543 set_velt_type(in, vt);
1544 }
1545 }
1546 }
1547 }
1548 }
1549 }
1550 #ifndef PRODUCT
1551 if (TraceSuperWord && Verbose) {
1552 for (int i = 0; i < _block.length(); i++) {
1553 Node* n = _block.at(i);
1554 velt_type(n)->dump();
1555 tty->print("\t");
1556 n->dump();
1557 }
1558 }
1559 #endif
1560 }
1561
1562 //------------------------------memory_alignment---------------------------
1563 // Alignment within a vector memory reference
1564 int SuperWord::memory_alignment(MemNode* s, int iv_adjust_in_bytes) {
1565 SWPointer p(s, this);
1566 if (!p.valid()) {
1567 return bottom_align;
1568 }
1569 int offset = p.offset_in_bytes();
1570 offset += iv_adjust_in_bytes;
1571 int off_rem = offset % vector_width_in_bytes();
1572 int off_mod = off_rem >= 0 ? off_rem : off_rem + vector_width_in_bytes();
1573 return off_mod;
1574 }
1575
1576 //---------------------------container_type---------------------------
1577 // Smallest type containing range of values
1578 const Type* SuperWord::container_type(const Type* t) {
1579 const Type* tp = t->make_ptr();
1580 if (tp && tp->isa_aryptr()) {
1581 t = tp->is_aryptr()->elem();
1582 }
1583 if (t->basic_type() == T_INT) {
1584 if (t->higher_equal(TypeInt::BOOL)) return TypeInt::BOOL;
1585 if (t->higher_equal(TypeInt::BYTE)) return TypeInt::BYTE;
1586 if (t->higher_equal(TypeInt::CHAR)) return TypeInt::CHAR;
1587 if (t->higher_equal(TypeInt::SHORT)) return TypeInt::SHORT;
1588 return TypeInt::INT;
1589 }
1590 return t;
1591 }
1592
1593 //-------------------------vector_opd_range-----------------------
1594 // (Start, end] half-open range defining which operands are vector
1595 void SuperWord::vector_opd_range(Node* n, uint* start, uint* end) {
1596 switch (n->Opcode()) {
1597 case Op_LoadB: case Op_LoadUS:
1598 case Op_LoadI: case Op_LoadL:
1599 case Op_LoadF: case Op_LoadD:
1600 case Op_LoadP:
1601 *start = 0;
1602 *end = 0;
1603 return;
1604 case Op_StoreB: case Op_StoreC:
1605 case Op_StoreI: case Op_StoreL:
1606 case Op_StoreF: case Op_StoreD:
1607 case Op_StoreP:
1608 *start = MemNode::ValueIn;
1609 *end = *start + 1;
1610 return;
1611 case Op_LShiftI: case Op_LShiftL:
1612 *start = 1;
1613 *end = 2;
1614 return;
1615 case Op_CMoveI: case Op_CMoveL: case Op_CMoveF: case Op_CMoveD:
1616 *start = 2;
1617 *end = n->req();
1618 return;
1619 }
1620 *start = 1;
1621 *end = n->req(); // default is all operands
1622 }
1623
1624 //------------------------------in_packset---------------------------
1625 // Are s1 and s2 in a pack pair and ordered as s1,s2?
1626 bool SuperWord::in_packset(Node* s1, Node* s2) {
1627 for (int i = 0; i < _packset.length(); i++) {
1628 Node_List* p = _packset.at(i);
1629 assert(p->size() == 2, "must be");
1630 if (p->at(0) == s1 && p->at(p->size()-1) == s2) {
1631 return true;
1632 }
1633 }
1634 return false;
1635 }
1636
1637 //------------------------------in_pack---------------------------
1638 // Is s in pack p?
1639 Node_List* SuperWord::in_pack(Node* s, Node_List* p) {
1640 for (uint i = 0; i < p->size(); i++) {
1641 if (p->at(i) == s) {
1642 return p;
1643 }
1644 }
1645 return NULL;
1646 }
1647
1648 //------------------------------remove_pack_at---------------------------
1649 // Remove the pack at position pos in the packset
1650 void SuperWord::remove_pack_at(int pos) {
1651 Node_List* p = _packset.at(pos);
1652 for (uint i = 0; i < p->size(); i++) {
1653 Node* s = p->at(i);
1654 set_my_pack(s, NULL);
1655 }
1656 _packset.remove_at(pos);
1657 }
1658
1659 //------------------------------executed_first---------------------------
1660 // Return the node executed first in pack p. Uses the RPO block list
1661 // to determine order.
1662 Node* SuperWord::executed_first(Node_List* p) {
1663 Node* n = p->at(0);
1664 int n_rpo = bb_idx(n);
1665 for (uint i = 1; i < p->size(); i++) {
1666 Node* s = p->at(i);
1667 int s_rpo = bb_idx(s);
1668 if (s_rpo < n_rpo) {
1669 n = s;
1670 n_rpo = s_rpo;
1671 }
1672 }
1673 return n;
1674 }
1675
1676 //------------------------------executed_last---------------------------
1677 // Return the node executed last in pack p.
1678 Node* SuperWord::executed_last(Node_List* p) {
1679 Node* n = p->at(0);
1680 int n_rpo = bb_idx(n);
1681 for (uint i = 1; i < p->size(); i++) {
1682 Node* s = p->at(i);
1683 int s_rpo = bb_idx(s);
1684 if (s_rpo > n_rpo) {
1685 n = s;
1686 n_rpo = s_rpo;
1687 }
1688 }
1689 return n;
1690 }
1691
1692 //----------------------------align_initial_loop_index---------------------------
1693 // Adjust pre-loop limit so that in main loop, a load/store reference
1694 // to align_to_ref will be a position zero in the vector.
1695 // (iv + k) mod vector_align == 0
1696 void SuperWord::align_initial_loop_index(MemNode* align_to_ref) {
1697 CountedLoopNode *main_head = lp()->as_CountedLoop();
1698 assert(main_head->is_main_loop(), "");
1699 CountedLoopEndNode* pre_end = get_pre_loop_end(main_head);
1700 assert(pre_end != NULL, "");
1701 Node *pre_opaq1 = pre_end->limit();
1702 assert(pre_opaq1->Opcode() == Op_Opaque1, "");
1703 Opaque1Node *pre_opaq = (Opaque1Node*)pre_opaq1;
1704 Node *lim0 = pre_opaq->in(1);
1705
1706 // Where we put new limit calculations
1707 Node *pre_ctrl = pre_end->loopnode()->in(LoopNode::EntryControl);
1708
1709 // Ensure the original loop limit is available from the
1710 // pre-loop Opaque1 node.
1711 Node *orig_limit = pre_opaq->original_loop_limit();
1712 assert(orig_limit != NULL && _igvn.type(orig_limit) != Type::TOP, "");
1713
1714 SWPointer align_to_ref_p(align_to_ref, this);
1715
1716 // Given:
1717 // lim0 == original pre loop limit
1718 // V == v_align (power of 2)
1719 // invar == extra invariant piece of the address expression
1720 // e == k [ +/- invar ]
1721 //
1722 // When reassociating expressions involving '%' the basic rules are:
1723 // (a - b) % k == 0 => a % k == b % k
1724 // and:
1725 // (a + b) % k == 0 => a % k == (k - b) % k
1726 //
1727 // For stride > 0 && scale > 0,
1728 // Derive the new pre-loop limit "lim" such that the two constraints:
1729 // (1) lim = lim0 + N (where N is some positive integer < V)
1730 // (2) (e + lim) % V == 0
1731 // are true.
1732 //
1733 // Substituting (1) into (2),
1734 // (e + lim0 + N) % V == 0
1735 // solve for N:
1736 // N = (V - (e + lim0)) % V
1737 // substitute back into (1), so that new limit
1738 // lim = lim0 + (V - (e + lim0)) % V
1739 //
1740 // For stride > 0 && scale < 0
1741 // Constraints:
1742 // lim = lim0 + N
1743 // (e - lim) % V == 0
1744 // Solving for lim:
1745 // (e - lim0 - N) % V == 0
1746 // N = (e - lim0) % V
1747 // lim = lim0 + (e - lim0) % V
1748 //
1749 // For stride < 0 && scale > 0
1750 // Constraints:
1751 // lim = lim0 - N
1752 // (e + lim) % V == 0
1753 // Solving for lim:
1754 // (e + lim0 - N) % V == 0
1755 // N = (e + lim0) % V
1756 // lim = lim0 - (e + lim0) % V
1757 //
1758 // For stride < 0 && scale < 0
1759 // Constraints:
1760 // lim = lim0 - N
1761 // (e - lim) % V == 0
1762 // Solving for lim:
1763 // (e - lim0 + N) % V == 0
1764 // N = (V - (e - lim0)) % V
1765 // lim = lim0 - (V - (e - lim0)) % V
1766
1767 int stride = iv_stride();
1768 int scale = align_to_ref_p.scale_in_bytes();
1769 int elt_size = align_to_ref_p.memory_size();
1770 int v_align = vector_width_in_bytes() / elt_size;
1771 int k = align_to_ref_p.offset_in_bytes() / elt_size;
1772
1773 Node *kn = _igvn.intcon(k);
1774
1775 Node *e = kn;
1776 if (align_to_ref_p.invar() != NULL) {
1777 // incorporate any extra invariant piece producing k +/- invar >>> log2(elt)
1778 Node* log2_elt = _igvn.intcon(exact_log2(elt_size));
1779 Node* aref = new (_phase->C, 3) URShiftINode(align_to_ref_p.invar(), log2_elt);
1780 _phase->_igvn.register_new_node_with_optimizer(aref);
1781 _phase->set_ctrl(aref, pre_ctrl);
1782 if (align_to_ref_p.negate_invar()) {
1783 e = new (_phase->C, 3) SubINode(e, aref);
1784 } else {
1785 e = new (_phase->C, 3) AddINode(e, aref);
1786 }
1787 _phase->_igvn.register_new_node_with_optimizer(e);
1788 _phase->set_ctrl(e, pre_ctrl);
1789 }
1790
1791 // compute e +/- lim0
1792 if (scale < 0) {
1793 e = new (_phase->C, 3) SubINode(e, lim0);
1794 } else {
1795 e = new (_phase->C, 3) AddINode(e, lim0);
1796 }
1797 _phase->_igvn.register_new_node_with_optimizer(e);
1798 _phase->set_ctrl(e, pre_ctrl);
1799
1800 if (stride * scale > 0) {
1801 // compute V - (e +/- lim0)
1802 Node* va = _igvn.intcon(v_align);
1803 e = new (_phase->C, 3) SubINode(va, e);
1804 _phase->_igvn.register_new_node_with_optimizer(e);
1805 _phase->set_ctrl(e, pre_ctrl);
1806 }
1807 // compute N = (exp) % V
1808 Node* va_msk = _igvn.intcon(v_align - 1);
1809 Node* N = new (_phase->C, 3) AndINode(e, va_msk);
1810 _phase->_igvn.register_new_node_with_optimizer(N);
1811 _phase->set_ctrl(N, pre_ctrl);
1812
1813 // substitute back into (1), so that new limit
1814 // lim = lim0 + N
1815 Node* lim;
1816 if (stride < 0) {
1817 lim = new (_phase->C, 3) SubINode(lim0, N);
1818 } else {
1819 lim = new (_phase->C, 3) AddINode(lim0, N);
1820 }
1821 _phase->_igvn.register_new_node_with_optimizer(lim);
1822 _phase->set_ctrl(lim, pre_ctrl);
1823 Node* constrained =
1824 (stride > 0) ? (Node*) new (_phase->C,3) MinINode(lim, orig_limit)
1825 : (Node*) new (_phase->C,3) MaxINode(lim, orig_limit);
1826 _phase->_igvn.register_new_node_with_optimizer(constrained);
1827 _phase->set_ctrl(constrained, pre_ctrl);
1828 _igvn.hash_delete(pre_opaq);
1829 pre_opaq->set_req(1, constrained);
1830 }
1831
1832 //----------------------------get_pre_loop_end---------------------------
1833 // Find pre loop end from main loop. Returns null if none.
1834 CountedLoopEndNode* SuperWord::get_pre_loop_end(CountedLoopNode *cl) {
1835 Node *ctrl = cl->in(LoopNode::EntryControl);
1836 if (!ctrl->is_IfTrue() && !ctrl->is_IfFalse()) return NULL;
1837 Node *iffm = ctrl->in(0);
1838 if (!iffm->is_If()) return NULL;
1839 Node *p_f = iffm->in(0);
1840 if (!p_f->is_IfFalse()) return NULL;
1841 if (!p_f->in(0)->is_CountedLoopEnd()) return NULL;
1842 CountedLoopEndNode *pre_end = p_f->in(0)->as_CountedLoopEnd();
1843 if (!pre_end->loopnode()->is_pre_loop()) return NULL;
1844 return pre_end;
1845 }
1846
1847
1848 //------------------------------init---------------------------
1849 void SuperWord::init() {
1850 _dg.init();
1851 _packset.clear();
1852 _disjoint_ptrs.clear();
1853 _block.clear();
1854 _data_entry.clear();
1855 _mem_slice_head.clear();
1856 _mem_slice_tail.clear();
1857 _node_info.clear();
1858 _align_to_ref = NULL;
1859 _lpt = NULL;
1860 _lp = NULL;
1861 _bb = NULL;
1862 _iv = NULL;
1863 }
1864
1865 //------------------------------print_packset---------------------------
1866 void SuperWord::print_packset() {
1867 #ifndef PRODUCT
1868 tty->print_cr("packset");
1869 for (int i = 0; i < _packset.length(); i++) {
1870 tty->print_cr("Pack: %d", i);
1871 Node_List* p = _packset.at(i);
1872 print_pack(p);
1873 }
1874 #endif
1875 }
1876
1877 //------------------------------print_pack---------------------------
1878 void SuperWord::print_pack(Node_List* p) {
1879 for (uint i = 0; i < p->size(); i++) {
1880 print_stmt(p->at(i));
1881 }
1882 }
1883
1884 //------------------------------print_bb---------------------------
1885 void SuperWord::print_bb() {
1886 #ifndef PRODUCT
1887 tty->print_cr("\nBlock");
1888 for (int i = 0; i < _block.length(); i++) {
1889 Node* n = _block.at(i);
1890 tty->print("%d ", i);
1891 if (n) {
1892 n->dump();
1893 }
1894 }
1895 #endif
1896 }
1897
1898 //------------------------------print_stmt---------------------------
1899 void SuperWord::print_stmt(Node* s) {
1900 #ifndef PRODUCT
1901 tty->print(" align: %d \t", alignment(s));
1902 s->dump();
1903 #endif
1904 }
1905
1906 //------------------------------blank---------------------------
1907 char* SuperWord::blank(uint depth) {
1908 static char blanks[101];
1909 assert(depth < 101, "too deep");
1910 for (uint i = 0; i < depth; i++) blanks[i] = ' ';
1911 blanks[depth] = '\0';
1912 return blanks;
1913 }
1914
1915
1916 //==============================SWPointer===========================
1917
1918 //----------------------------SWPointer------------------------
1919 SWPointer::SWPointer(MemNode* mem, SuperWord* slp) :
1920 _mem(mem), _slp(slp), _base(NULL), _adr(NULL),
1921 _scale(0), _offset(0), _invar(NULL), _negate_invar(false) {
1922
1923 Node* adr = mem->in(MemNode::Address);
1924 if (!adr->is_AddP()) {
1925 assert(!valid(), "too complex");
1926 return;
1927 }
1928 // Match AddP(base, AddP(ptr, k*iv [+ invariant]), constant)
1929 Node* base = adr->in(AddPNode::Base);
1930 //unsafe reference could not be aligned appropriately without runtime checking
1931 if (base == NULL || base->bottom_type() == Type::TOP) {
1932 assert(!valid(), "unsafe access");
1933 return;
1934 }
1935 for (int i = 0; i < 3; i++) {
1936 if (!scaled_iv_plus_offset(adr->in(AddPNode::Offset))) {
1937 assert(!valid(), "too complex");
1938 return;
1939 }
1940 adr = adr->in(AddPNode::Address);
1941 if (base == adr || !adr->is_AddP()) {
1942 break; // stop looking at addp's
1943 }
1944 }
1945 _base = base;
1946 _adr = adr;
1947 assert(valid(), "Usable");
1948 }
1949
1950 // Following is used to create a temporary object during
1951 // the pattern match of an address expression.
1952 SWPointer::SWPointer(SWPointer* p) :
1953 _mem(p->_mem), _slp(p->_slp), _base(NULL), _adr(NULL),
1954 _scale(0), _offset(0), _invar(NULL), _negate_invar(false) {}
1955
1956 //------------------------scaled_iv_plus_offset--------------------
1957 // Match: k*iv + offset
1958 // where: k is a constant that maybe zero, and
1959 // offset is (k2 [+/- invariant]) where k2 maybe zero and invariant is optional
1960 bool SWPointer::scaled_iv_plus_offset(Node* n) {
1961 if (scaled_iv(n)) {
1962 return true;
1963 }
1964 if (offset_plus_k(n)) {
1965 return true;
1966 }
1967 int opc = n->Opcode();
1968 if (opc == Op_AddI) {
1969 if (scaled_iv(n->in(1)) && offset_plus_k(n->in(2))) {
1970 return true;
1971 }
1972 if (scaled_iv(n->in(2)) && offset_plus_k(n->in(1))) {
1973 return true;
1974 }
1975 } else if (opc == Op_SubI) {
1976 if (scaled_iv(n->in(1)) && offset_plus_k(n->in(2), true)) {
1977 return true;
1978 }
1979 if (scaled_iv(n->in(2)) && offset_plus_k(n->in(1))) {
1980 _scale *= -1;
1981 return true;
1982 }
1983 }
1984 return false;
1985 }
1986
1987 //----------------------------scaled_iv------------------------
1988 // Match: k*iv where k is a constant that's not zero
1989 bool SWPointer::scaled_iv(Node* n) {
1990 if (_scale != 0) {
1991 return false; // already found a scale
1992 }
1993 if (n == iv()) {
1994 _scale = 1;
1995 return true;
1996 }
1997 int opc = n->Opcode();
1998 if (opc == Op_MulI) {
1999 if (n->in(1) == iv() && n->in(2)->is_Con()) {
2000 _scale = n->in(2)->get_int();
2001 return true;
2002 } else if (n->in(2) == iv() && n->in(1)->is_Con()) {
2003 _scale = n->in(1)->get_int();
2004 return true;
2005 }
2006 } else if (opc == Op_LShiftI) {
2007 if (n->in(1) == iv() && n->in(2)->is_Con()) {
2008 _scale = 1 << n->in(2)->get_int();
2009 return true;
2010 }
2011 } else if (opc == Op_ConvI2L) {
2012 if (scaled_iv_plus_offset(n->in(1))) {
2013 return true;
2014 }
2015 } else if (opc == Op_LShiftL) {
2016 if (!has_iv() && _invar == NULL) {
2017 // Need to preserve the current _offset value, so
2018 // create a temporary object for this expression subtree.
2019 // Hacky, so should re-engineer the address pattern match.
2020 SWPointer tmp(this);
2021 if (tmp.scaled_iv_plus_offset(n->in(1))) {
2022 if (tmp._invar == NULL) {
2023 int mult = 1 << n->in(2)->get_int();
2024 _scale = tmp._scale * mult;
2025 _offset += tmp._offset * mult;
2026 return true;
2027 }
2028 }
2029 }
2030 }
2031 return false;
2032 }
2033
2034 //----------------------------offset_plus_k------------------------
2035 // Match: offset is (k [+/- invariant])
2036 // where k maybe zero and invariant is optional, but not both.
2037 bool SWPointer::offset_plus_k(Node* n, bool negate) {
2038 int opc = n->Opcode();
2039 if (opc == Op_ConI) {
2040 _offset += negate ? -(n->get_int()) : n->get_int();
2041 return true;
2042 } else if (opc == Op_ConL) {
2043 // Okay if value fits into an int
2044 const TypeLong* t = n->find_long_type();
2045 if (t->higher_equal(TypeLong::INT)) {
2046 jlong loff = n->get_long();
2047 jint off = (jint)loff;
2048 _offset += negate ? -off : loff;
2049 return true;
2050 }
2051 return false;
2052 }
2053 if (_invar != NULL) return false; // already have an invariant
2054 if (opc == Op_AddI) {
2055 if (n->in(2)->is_Con() && invariant(n->in(1))) {
2056 _negate_invar = negate;
2057 _invar = n->in(1);
2058 _offset += negate ? -(n->in(2)->get_int()) : n->in(2)->get_int();
2059 return true;
2060 } else if (n->in(1)->is_Con() && invariant(n->in(2))) {
2061 _offset += negate ? -(n->in(1)->get_int()) : n->in(1)->get_int();
2062 _negate_invar = negate;
2063 _invar = n->in(2);
2064 return true;
2065 }
2066 }
2067 if (opc == Op_SubI) {
2068 if (n->in(2)->is_Con() && invariant(n->in(1))) {
2069 _negate_invar = negate;
2070 _invar = n->in(1);
2071 _offset += !negate ? -(n->in(2)->get_int()) : n->in(2)->get_int();
2072 return true;
2073 } else if (n->in(1)->is_Con() && invariant(n->in(2))) {
2074 _offset += negate ? -(n->in(1)->get_int()) : n->in(1)->get_int();
2075 _negate_invar = !negate;
2076 _invar = n->in(2);
2077 return true;
2078 }
2079 }
2080 if (invariant(n)) {
2081 _negate_invar = negate;
2082 _invar = n;
2083 return true;
2084 }
2085 return false;
2086 }
2087
2088 //----------------------------print------------------------
2089 void SWPointer::print() {
2090 #ifndef PRODUCT
2091 tty->print("base: %d adr: %d scale: %d offset: %d invar: %c%d\n",
2092 _base != NULL ? _base->_idx : 0,
2093 _adr != NULL ? _adr->_idx : 0,
2094 _scale, _offset,
2095 _negate_invar?'-':'+',
2096 _invar != NULL ? _invar->_idx : 0);
2097 #endif
2098 }
2099
2100 // ========================= OrderedPair =====================
2101
2102 const OrderedPair OrderedPair::initial;
2103
2104 // ========================= SWNodeInfo =====================
2105
2106 const SWNodeInfo SWNodeInfo::initial;
2107
2108
2109 // ============================ DepGraph ===========================
2110
2111 //------------------------------make_node---------------------------
2112 // Make a new dependence graph node for an ideal node.
2113 DepMem* DepGraph::make_node(Node* node) {
2114 DepMem* m = new (_arena) DepMem(node);
2115 if (node != NULL) {
2116 assert(_map.at_grow(node->_idx) == NULL, "one init only");
2117 _map.at_put_grow(node->_idx, m);
2118 }
2119 return m;
2120 }
2121
2122 //------------------------------make_edge---------------------------
2123 // Make a new dependence graph edge from dpred -> dsucc
2124 DepEdge* DepGraph::make_edge(DepMem* dpred, DepMem* dsucc) {
2125 DepEdge* e = new (_arena) DepEdge(dpred, dsucc, dsucc->in_head(), dpred->out_head());
2126 dpred->set_out_head(e);
2127 dsucc->set_in_head(e);
2128 return e;
2129 }
2130
2131 // ========================== DepMem ========================
2132
2133 //------------------------------in_cnt---------------------------
2134 int DepMem::in_cnt() {
2135 int ct = 0;
2136 for (DepEdge* e = _in_head; e != NULL; e = e->next_in()) ct++;
2137 return ct;
2138 }
2139
2140 //------------------------------out_cnt---------------------------
2141 int DepMem::out_cnt() {
2142 int ct = 0;
2143 for (DepEdge* e = _out_head; e != NULL; e = e->next_out()) ct++;
2144 return ct;
2145 }
2146
2147 //------------------------------print-----------------------------
2148 void DepMem::print() {
2149 #ifndef PRODUCT
2150 tty->print(" DepNode %d (", _node->_idx);
2151 for (DepEdge* p = _in_head; p != NULL; p = p->next_in()) {
2152 Node* pred = p->pred()->node();
2153 tty->print(" %d", pred != NULL ? pred->_idx : 0);
2154 }
2155 tty->print(") [");
2156 for (DepEdge* s = _out_head; s != NULL; s = s->next_out()) {
2157 Node* succ = s->succ()->node();
2158 tty->print(" %d", succ != NULL ? succ->_idx : 0);
2159 }
2160 tty->print_cr(" ]");
2161 #endif
2162 }
2163
2164 // =========================== DepEdge =========================
2165
2166 //------------------------------DepPreds---------------------------
2167 void DepEdge::print() {
2168 #ifndef PRODUCT
2169 tty->print_cr("DepEdge: %d [ %d ]", _pred->node()->_idx, _succ->node()->_idx);
2170 #endif
2171 }
2172
2173 // =========================== DepPreds =========================
2174 // Iterator over predecessor edges in the dependence graph.
2175
2176 //------------------------------DepPreds---------------------------
2177 DepPreds::DepPreds(Node* n, DepGraph& dg) {
2178 _n = n;
2179 _done = false;
2180 if (_n->is_Store() || _n->is_Load()) {
2181 _next_idx = MemNode::Address;
2182 _end_idx = n->req();
2183 _dep_next = dg.dep(_n)->in_head();
2184 } else if (_n->is_Mem()) {
2185 _next_idx = 0;
2186 _end_idx = 0;
2187 _dep_next = dg.dep(_n)->in_head();
2188 } else {
2189 _next_idx = 1;
2190 _end_idx = _n->req();
2191 _dep_next = NULL;
2192 }
2193 next();
2194 }
2195
2196 //------------------------------next---------------------------
2197 void DepPreds::next() {
2198 if (_dep_next != NULL) {
2199 _current = _dep_next->pred()->node();
2200 _dep_next = _dep_next->next_in();
2201 } else if (_next_idx < _end_idx) {
2202 _current = _n->in(_next_idx++);
2203 } else {
2204 _done = true;
2205 }
2206 }
2207
2208 // =========================== DepSuccs =========================
2209 // Iterator over successor edges in the dependence graph.
2210
2211 //------------------------------DepSuccs---------------------------
2212 DepSuccs::DepSuccs(Node* n, DepGraph& dg) {
2213 _n = n;
2214 _done = false;
2215 if (_n->is_Load()) {
2216 _next_idx = 0;
2217 _end_idx = _n->outcnt();
2218 _dep_next = dg.dep(_n)->out_head();
2219 } else if (_n->is_Mem() || _n->is_Phi() && _n->bottom_type() == Type::MEMORY) {
2220 _next_idx = 0;
2221 _end_idx = 0;
2222 _dep_next = dg.dep(_n)->out_head();
2223 } else {
2224 _next_idx = 0;
2225 _end_idx = _n->outcnt();
2226 _dep_next = NULL;
2227 }
2228 next();
2229 }
2230
2231 //-------------------------------next---------------------------
2232 void DepSuccs::next() {
2233 if (_dep_next != NULL) {
2234 _current = _dep_next->succ()->node();
2235 _dep_next = _dep_next->next_out();
2236 } else if (_next_idx < _end_idx) {
2237 _current = _n->raw_out(_next_idx++);
2238 } else {
2239 _done = true;
2240 }
2241 }