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 "precompiled.hpp"
25 #include "compiler/compileLog.hpp"
26 #include "libadt/vectset.hpp"
27 #include "memory/allocation.inline.hpp"
28 #include "opto/addnode.hpp"
29 #include "opto/callnode.hpp"
30 #include "opto/divnode.hpp"
31 #include "opto/matcher.hpp"
32 #include "opto/memnode.hpp"
33 #include "opto/mulnode.hpp"
34 #include "opto/opcodes.hpp"
35 #include "opto/superword.hpp"
36 #include "opto/vectornode.hpp"
37
38 //
39 // S U P E R W O R D T R A N S F O R M
40 //=============================================================================
41
42 //------------------------------SuperWord---------------------------
43 SuperWord::SuperWord(PhaseIdealLoop* phase) :
44 _phase(phase),
45 _igvn(phase->_igvn),
46 _arena(phase->C->comp_arena()),
47 _packset(arena(), 8, 0, NULL), // packs for the current block
48 _bb_idx(arena(), (int)(1.10 * phase->C->unique()), 0, 0), // node idx to index in bb
49 _block(arena(), 8, 0, NULL), // nodes in current block
50 _data_entry(arena(), 8, 0, NULL), // nodes with all inputs from outside
51 _mem_slice_head(arena(), 8, 0, NULL), // memory slice heads
52 _mem_slice_tail(arena(), 8, 0, NULL), // memory slice tails
53 _node_info(arena(), 8, 0, SWNodeInfo::initial), // info needed per node
54 _align_to_ref(NULL), // memory reference to align vectors to
55 _disjoint_ptrs(arena(), 8, 0, OrderedPair::initial), // runtime disambiguated pointer pairs
56 _dg(_arena), // dependence graph
57 _visited(arena()), // visited node set
58 _post_visited(arena()), // post visited node set
59 _n_idx_list(arena(), 8), // scratch list of (node,index) pairs
60 _stk(arena(), 8, 0, NULL), // scratch stack of nodes
61 _nlist(arena(), 8, 0, NULL), // scratch list of nodes
62 _lpt(NULL), // loop tree node
63 _lp(NULL), // LoopNode
64 _bb(NULL), // basic block
65 _iv(NULL) // induction var
66 {}
67
68 //------------------------------transform_loop---------------------------
69 void SuperWord::transform_loop(IdealLoopTree* lpt) {
70 assert(lpt->_head->is_CountedLoop(), "must be");
71 CountedLoopNode *cl = lpt->_head->as_CountedLoop();
72
73 if (!cl->is_main_loop() ) return; // skip normal, pre, and post loops
74
75 // Check for no control flow in body (other than exit)
76 Node *cl_exit = cl->loopexit();
77 if (cl_exit->in(0) != lpt->_head) return;
78
79 // Make sure the are no extra control users of the loop backedge
80 if (cl->back_control()->outcnt() != 1) {
81 return;
82 }
83
84 // Check for pre-loop ending with CountedLoopEnd(Bool(Cmp(x,Opaque1(limit))))
85 CountedLoopEndNode* pre_end = get_pre_loop_end(cl);
86 if (pre_end == NULL) return;
87 Node *pre_opaq1 = pre_end->limit();
88 if (pre_opaq1->Opcode() != Op_Opaque1) return;
89
90 // Do vectors exist on this architecture?
91 if (vector_width_in_bytes() == 0) return;
92
93 init(); // initialize data structures
94
95 set_lpt(lpt);
96 set_lp(cl);
97
98 // For now, define one block which is the entire loop body
99 set_bb(cl);
100
101 assert(_packset.length() == 0, "packset must be empty");
102 SLP_extract();
103 }
104
105 //------------------------------SLP_extract---------------------------
106 // Extract the superword level parallelism
107 //
108 // 1) A reverse post-order of nodes in the block is constructed. By scanning
109 // this list from first to last, all definitions are visited before their uses.
110 //
111 // 2) A point-to-point dependence graph is constructed between memory references.
112 // This simplies the upcoming "independence" checker.
113 //
114 // 3) The maximum depth in the node graph from the beginning of the block
115 // to each node is computed. This is used to prune the graph search
116 // in the independence checker.
117 //
118 // 4) For integer types, the necessary bit width is propagated backwards
119 // from stores to allow packed operations on byte, char, and short
120 // integers. This reverses the promotion to type "int" that javac
121 // did for operations like: char c1,c2,c3; c1 = c2 + c3.
122 //
123 // 5) One of the memory references is picked to be an aligned vector reference.
124 // The pre-loop trip count is adjusted to align this reference in the
125 // unrolled body.
126 //
127 // 6) The initial set of pack pairs is seeded with memory references.
128 //
129 // 7) The set of pack pairs is extended by following use->def and def->use links.
130 //
131 // 8) The pairs are combined into vector sized packs.
132 //
133 // 9) Reorder the memory slices to co-locate members of the memory packs.
134 //
135 // 10) Generate ideal vector nodes for the final set of packs and where necessary,
136 // inserting scalar promotion, vector creation from multiple scalars, and
137 // extraction of scalar values from vectors.
138 //
139 void SuperWord::SLP_extract() {
140
141 // Ready the block
142
143 construct_bb();
144
145 dependence_graph();
146
147 compute_max_depth();
148
149 compute_vector_element_type();
150
151 // Attempt vectorization
152
153 find_adjacent_refs();
154
155 extend_packlist();
156
157 combine_packs();
158
159 construct_my_pack_map();
160
161 filter_packs();
162
163 schedule();
164
165 output();
166 }
167
168 //------------------------------find_adjacent_refs---------------------------
169 // Find the adjacent memory references and create pack pairs for them.
170 // This is the initial set of packs that will then be extended by
171 // following use->def and def->use links. The align positions are
172 // assigned relative to the reference "align_to_ref"
173 void SuperWord::find_adjacent_refs() {
174 // Get list of memory operations
175 Node_List memops;
176 for (int i = 0; i < _block.length(); i++) {
177 Node* n = _block.at(i);
178 if (n->is_Mem() && in_bb(n) &&
179 is_java_primitive(n->as_Mem()->memory_type())) {
180 int align = memory_alignment(n->as_Mem(), 0);
181 if (align != bottom_align) {
182 memops.push(n);
183 }
184 }
185 }
186 if (memops.size() == 0) return;
187
188 // Find a memory reference to align to. The pre-loop trip count
189 // is modified to align this reference to a vector-aligned address
190 find_align_to_ref(memops);
191 if (align_to_ref() == NULL) return;
192
193 SWPointer align_to_ref_p(align_to_ref(), this);
194 int offset = align_to_ref_p.offset_in_bytes();
195 int scale = align_to_ref_p.scale_in_bytes();
196 int vw = vector_width_in_bytes();
197 int stride_sign = (scale * iv_stride()) > 0 ? 1 : -1;
198 int iv_adjustment = (stride_sign * vw - (offset % vw)) % vw;
199
200 #ifndef PRODUCT
201 if (TraceSuperWord)
202 tty->print_cr("\noffset = %d iv_adjustment = %d elt_align = %d scale = %d iv_stride = %d",
203 offset, iv_adjustment, align_to_ref_p.memory_size(), align_to_ref_p.scale_in_bytes(), iv_stride());
204 #endif
205
206 // Set alignment relative to "align_to_ref"
207 for (int i = memops.size() - 1; i >= 0; i--) {
208 MemNode* s = memops.at(i)->as_Mem();
209 SWPointer p2(s, this);
210 if (p2.comparable(align_to_ref_p)) {
211 int align = memory_alignment(s, iv_adjustment);
212 set_alignment(s, align);
213 } else {
214 memops.remove(i);
215 }
216 }
217
218 // Create initial pack pairs of memory operations
219 for (uint i = 0; i < memops.size(); i++) {
220 Node* s1 = memops.at(i);
221 for (uint j = 0; j < memops.size(); j++) {
222 Node* s2 = memops.at(j);
223 if (s1 != s2 && are_adjacent_refs(s1, s2)) {
224 int align = alignment(s1);
225 if (stmts_can_pack(s1, s2, align)) {
226 Node_List* pair = new Node_List();
227 pair->push(s1);
228 pair->push(s2);
229 _packset.append(pair);
230 }
231 }
232 }
233 }
234
235 #ifndef PRODUCT
236 if (TraceSuperWord) {
237 tty->print_cr("\nAfter find_adjacent_refs");
238 print_packset();
239 }
240 #endif
241 }
242
243 //------------------------------find_align_to_ref---------------------------
244 // Find a memory reference to align the loop induction variable to.
245 // Looks first at stores then at loads, looking for a memory reference
246 // with the largest number of references similar to it.
247 void SuperWord::find_align_to_ref(Node_List &memops) {
248 GrowableArray<int> cmp_ct(arena(), memops.size(), memops.size(), 0);
249
250 // Count number of comparable memory ops
251 for (uint i = 0; i < memops.size(); i++) {
252 MemNode* s1 = memops.at(i)->as_Mem();
253 SWPointer p1(s1, this);
254 // Discard if pre loop can't align this reference
255 if (!ref_is_alignable(p1)) {
256 *cmp_ct.adr_at(i) = 0;
257 continue;
258 }
259 for (uint j = i+1; j < memops.size(); j++) {
260 MemNode* s2 = memops.at(j)->as_Mem();
261 if (isomorphic(s1, s2)) {
262 SWPointer p2(s2, this);
263 if (p1.comparable(p2)) {
264 (*cmp_ct.adr_at(i))++;
265 (*cmp_ct.adr_at(j))++;
266 }
267 }
268 }
269 }
270
271 // Find Store (or Load) with the greatest number of "comparable" references
272 int max_ct = 0;
273 int max_idx = -1;
274 int min_size = max_jint;
275 int min_iv_offset = max_jint;
276 for (uint j = 0; j < memops.size(); j++) {
277 MemNode* s = memops.at(j)->as_Mem();
278 if (s->is_Store()) {
279 SWPointer p(s, this);
280 if (cmp_ct.at(j) > max_ct ||
281 cmp_ct.at(j) == max_ct && (data_size(s) < min_size ||
282 data_size(s) == min_size &&
283 p.offset_in_bytes() < min_iv_offset)) {
284 max_ct = cmp_ct.at(j);
285 max_idx = j;
286 min_size = data_size(s);
287 min_iv_offset = p.offset_in_bytes();
288 }
289 }
290 }
291 // If no stores, look at loads
292 if (max_ct == 0) {
293 for (uint j = 0; j < memops.size(); j++) {
294 MemNode* s = memops.at(j)->as_Mem();
295 if (s->is_Load()) {
296 SWPointer p(s, this);
297 if (cmp_ct.at(j) > max_ct ||
298 cmp_ct.at(j) == max_ct && (data_size(s) < min_size ||
299 data_size(s) == min_size &&
300 p.offset_in_bytes() < min_iv_offset)) {
301 max_ct = cmp_ct.at(j);
302 max_idx = j;
303 min_size = data_size(s);
304 min_iv_offset = p.offset_in_bytes();
305 }
306 }
307 }
308 }
309
310 if (max_ct > 0)
311 set_align_to_ref(memops.at(max_idx)->as_Mem());
312
313 #ifndef PRODUCT
314 if (TraceSuperWord && Verbose) {
315 tty->print_cr("\nVector memops after find_align_to_refs");
316 for (uint i = 0; i < memops.size(); i++) {
317 MemNode* s = memops.at(i)->as_Mem();
318 s->dump();
319 }
320 }
321 #endif
322 }
323
324 //------------------------------ref_is_alignable---------------------------
325 // Can the preloop align the reference to position zero in the vector?
326 bool SuperWord::ref_is_alignable(SWPointer& p) {
327 if (!p.has_iv()) {
328 return true; // no induction variable
329 }
330 CountedLoopEndNode* pre_end = get_pre_loop_end(lp()->as_CountedLoop());
331 assert(pre_end->stride_is_con(), "pre loop stride is constant");
332 int preloop_stride = pre_end->stride_con();
333
334 int span = preloop_stride * p.scale_in_bytes();
335
336 // Stride one accesses are alignable.
337 if (ABS(span) == p.memory_size())
338 return true;
339
340 // If initial offset from start of object is computable,
341 // compute alignment within the vector.
342 int vw = vector_width_in_bytes();
343 if (vw % span == 0) {
344 Node* init_nd = pre_end->init_trip();
345 if (init_nd->is_Con() && p.invar() == NULL) {
346 int init = init_nd->bottom_type()->is_int()->get_con();
347
348 int init_offset = init * p.scale_in_bytes() + p.offset_in_bytes();
349 assert(init_offset >= 0, "positive offset from object start");
350
351 if (span > 0) {
352 return (vw - (init_offset % vw)) % span == 0;
353 } else {
354 assert(span < 0, "nonzero stride * scale");
355 return (init_offset % vw) % -span == 0;
356 }
357 }
358 }
359 return false;
360 }
361
362 //---------------------------dependence_graph---------------------------
363 // Construct dependency graph.
364 // Add dependence edges to load/store nodes for memory dependence
365 // A.out()->DependNode.in(1) and DependNode.out()->B.prec(x)
366 void SuperWord::dependence_graph() {
367 // First, assign a dependence node to each memory node
368 for (int i = 0; i < _block.length(); i++ ) {
369 Node *n = _block.at(i);
370 if (n->is_Mem() || n->is_Phi() && n->bottom_type() == Type::MEMORY) {
371 _dg.make_node(n);
372 }
373 }
374
375 // For each memory slice, create the dependences
376 for (int i = 0; i < _mem_slice_head.length(); i++) {
377 Node* n = _mem_slice_head.at(i);
378 Node* n_tail = _mem_slice_tail.at(i);
379
380 // Get slice in predecessor order (last is first)
381 mem_slice_preds(n_tail, n, _nlist);
382
383 // Make the slice dependent on the root
384 DepMem* slice = _dg.dep(n);
385 _dg.make_edge(_dg.root(), slice);
386
387 // Create a sink for the slice
388 DepMem* slice_sink = _dg.make_node(NULL);
389 _dg.make_edge(slice_sink, _dg.tail());
390
391 // Now visit each pair of memory ops, creating the edges
392 for (int j = _nlist.length() - 1; j >= 0 ; j--) {
393 Node* s1 = _nlist.at(j);
394
395 // If no dependency yet, use slice
396 if (_dg.dep(s1)->in_cnt() == 0) {
397 _dg.make_edge(slice, s1);
398 }
399 SWPointer p1(s1->as_Mem(), this);
400 bool sink_dependent = true;
401 for (int k = j - 1; k >= 0; k--) {
402 Node* s2 = _nlist.at(k);
403 if (s1->is_Load() && s2->is_Load())
404 continue;
405 SWPointer p2(s2->as_Mem(), this);
406
407 int cmp = p1.cmp(p2);
408 if (SuperWordRTDepCheck &&
409 p1.base() != p2.base() && p1.valid() && p2.valid()) {
410 // Create a runtime check to disambiguate
411 OrderedPair pp(p1.base(), p2.base());
412 _disjoint_ptrs.append_if_missing(pp);
413 } else if (!SWPointer::not_equal(cmp)) {
414 // Possibly same address
415 _dg.make_edge(s1, s2);
416 sink_dependent = false;
417 }
418 }
419 if (sink_dependent) {
420 _dg.make_edge(s1, slice_sink);
421 }
422 }
423 #ifndef PRODUCT
424 if (TraceSuperWord) {
425 tty->print_cr("\nDependence graph for slice: %d", n->_idx);
426 for (int q = 0; q < _nlist.length(); q++) {
427 _dg.print(_nlist.at(q));
428 }
429 tty->cr();
430 }
431 #endif
432 _nlist.clear();
433 }
434
435 #ifndef PRODUCT
436 if (TraceSuperWord) {
437 tty->print_cr("\ndisjoint_ptrs: %s", _disjoint_ptrs.length() > 0 ? "" : "NONE");
438 for (int r = 0; r < _disjoint_ptrs.length(); r++) {
439 _disjoint_ptrs.at(r).print();
440 tty->cr();
441 }
442 tty->cr();
443 }
444 #endif
445 }
446
447 //---------------------------mem_slice_preds---------------------------
448 // Return a memory slice (node list) in predecessor order starting at "start"
449 void SuperWord::mem_slice_preds(Node* start, Node* stop, GrowableArray<Node*> &preds) {
450 assert(preds.length() == 0, "start empty");
451 Node* n = start;
452 Node* prev = NULL;
453 while (true) {
454 assert(in_bb(n), "must be in block");
455 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
456 Node* out = n->fast_out(i);
457 if (out->is_Load()) {
458 if (in_bb(out)) {
459 preds.push(out);
460 }
461 } else {
462 // FIXME
463 if (out->is_MergeMem() && !in_bb(out)) {
464 // Either unrolling is causing a memory edge not to disappear,
465 // or need to run igvn.optimize() again before SLP
466 } else if (out->is_Phi() && out->bottom_type() == Type::MEMORY && !in_bb(out)) {
467 // Ditto. Not sure what else to check further.
468 } else if (out->Opcode() == Op_StoreCM && out->in(MemNode::OopStore) == n) {
469 // StoreCM has an input edge used as a precedence edge.
470 // Maybe an issue when oop stores are vectorized.
471 } else {
472 assert(out == prev || prev == NULL, "no branches off of store slice");
473 }
474 }
475 }
476 if (n == stop) break;
477 preds.push(n);
478 prev = n;
479 n = n->in(MemNode::Memory);
480 }
481 }
482
483 //------------------------------stmts_can_pack---------------------------
484 // Can s1 and s2 be in a pack with s1 immediately preceding s2 and
485 // s1 aligned at "align"
486 bool SuperWord::stmts_can_pack(Node* s1, Node* s2, int align) {
487
488 // Do not use superword for non-primitives
489 if((s1->is_Mem() && !is_java_primitive(s1->as_Mem()->memory_type())) ||
490 (s2->is_Mem() && !is_java_primitive(s2->as_Mem()->memory_type())))
491 return false;
492
493 if (isomorphic(s1, s2)) {
494 if (independent(s1, s2)) {
495 if (!exists_at(s1, 0) && !exists_at(s2, 1)) {
496 if (!s1->is_Mem() || are_adjacent_refs(s1, s2)) {
497 int s1_align = alignment(s1);
498 int s2_align = alignment(s2);
499 if (s1_align == top_align || s1_align == align) {
500 if (s2_align == top_align || s2_align == align + data_size(s1)) {
501 return true;
502 }
503 }
504 }
505 }
506 }
507 }
508 return false;
509 }
510
511 //------------------------------exists_at---------------------------
512 // Does s exist in a pack at position pos?
513 bool SuperWord::exists_at(Node* s, uint pos) {
514 for (int i = 0; i < _packset.length(); i++) {
515 Node_List* p = _packset.at(i);
516 if (p->at(pos) == s) {
517 return true;
518 }
519 }
520 return false;
521 }
522
523 //------------------------------are_adjacent_refs---------------------------
524 // Is s1 immediately before s2 in memory?
525 bool SuperWord::are_adjacent_refs(Node* s1, Node* s2) {
526 if (!s1->is_Mem() || !s2->is_Mem()) return false;
527 if (!in_bb(s1) || !in_bb(s2)) return false;
528
529 // Do not use superword for non-primitives
530 if (!is_java_primitive(s1->as_Mem()->memory_type()) ||
531 !is_java_primitive(s2->as_Mem()->memory_type())) {
532 return false;
533 }
534
535 // FIXME - co_locate_pack fails on Stores in different mem-slices, so
536 // only pack memops that are in the same alias set until that's fixed.
537 if (_phase->C->get_alias_index(s1->as_Mem()->adr_type()) !=
538 _phase->C->get_alias_index(s2->as_Mem()->adr_type()))
539 return false;
540 SWPointer p1(s1->as_Mem(), this);
541 SWPointer p2(s2->as_Mem(), this);
542 if (p1.base() != p2.base() || !p1.comparable(p2)) return false;
543 int diff = p2.offset_in_bytes() - p1.offset_in_bytes();
544 return diff == data_size(s1);
545 }
546
547 //------------------------------isomorphic---------------------------
548 // Are s1 and s2 similar?
549 bool SuperWord::isomorphic(Node* s1, Node* s2) {
550 if (s1->Opcode() != s2->Opcode()) return false;
551 if (s1->req() != s2->req()) return false;
552 if (s1->in(0) != s2->in(0)) return false;
553 if (velt_type(s1) != velt_type(s2)) return false;
554 return true;
555 }
556
557 //------------------------------independent---------------------------
558 // Is there no data path from s1 to s2 or s2 to s1?
559 bool SuperWord::independent(Node* s1, Node* s2) {
560 // assert(s1->Opcode() == s2->Opcode(), "check isomorphic first");
561 int d1 = depth(s1);
562 int d2 = depth(s2);
563 if (d1 == d2) return s1 != s2;
564 Node* deep = d1 > d2 ? s1 : s2;
565 Node* shallow = d1 > d2 ? s2 : s1;
566
567 visited_clear();
568
569 return independent_path(shallow, deep);
570 }
571
572 //------------------------------independent_path------------------------------
573 // Helper for independent
574 bool SuperWord::independent_path(Node* shallow, Node* deep, uint dp) {
575 if (dp >= 1000) return false; // stop deep recursion
576 visited_set(deep);
577 int shal_depth = depth(shallow);
578 assert(shal_depth <= depth(deep), "must be");
579 for (DepPreds preds(deep, _dg); !preds.done(); preds.next()) {
580 Node* pred = preds.current();
581 if (in_bb(pred) && !visited_test(pred)) {
582 if (shallow == pred) {
583 return false;
584 }
585 if (shal_depth < depth(pred) && !independent_path(shallow, pred, dp+1)) {
586 return false;
587 }
588 }
589 }
590 return true;
591 }
592
593 //------------------------------set_alignment---------------------------
594 void SuperWord::set_alignment(Node* s1, Node* s2, int align) {
595 set_alignment(s1, align);
596 set_alignment(s2, align + data_size(s1));
597 }
598
599 //------------------------------data_size---------------------------
600 int SuperWord::data_size(Node* s) {
601 const Type* t = velt_type(s);
602 BasicType bt = t->array_element_basic_type();
603 int bsize = type2aelembytes(bt);
604 assert(bsize != 0, "valid size");
605 return bsize;
606 }
607
608 //------------------------------extend_packlist---------------------------
609 // Extend packset by following use->def and def->use links from pack members.
610 void SuperWord::extend_packlist() {
611 bool changed;
612 do {
613 changed = false;
614 for (int i = 0; i < _packset.length(); i++) {
615 Node_List* p = _packset.at(i);
616 changed |= follow_use_defs(p);
617 changed |= follow_def_uses(p);
618 }
619 } while (changed);
620
621 #ifndef PRODUCT
622 if (TraceSuperWord) {
623 tty->print_cr("\nAfter extend_packlist");
624 print_packset();
625 }
626 #endif
627 }
628
629 //------------------------------follow_use_defs---------------------------
630 // Extend the packset by visiting operand definitions of nodes in pack p
631 bool SuperWord::follow_use_defs(Node_List* p) {
632 Node* s1 = p->at(0);
633 Node* s2 = p->at(1);
634 assert(p->size() == 2, "just checking");
635 assert(s1->req() == s2->req(), "just checking");
636 assert(alignment(s1) + data_size(s1) == alignment(s2), "just checking");
637
638 if (s1->is_Load()) return false;
639
640 int align = alignment(s1);
641 bool changed = false;
642 int start = s1->is_Store() ? MemNode::ValueIn : 1;
643 int end = s1->is_Store() ? MemNode::ValueIn+1 : s1->req();
644 for (int j = start; j < end; j++) {
645 Node* t1 = s1->in(j);
646 Node* t2 = s2->in(j);
647 if (!in_bb(t1) || !in_bb(t2))
648 continue;
649 if (stmts_can_pack(t1, t2, align)) {
650 if (est_savings(t1, t2) >= 0) {
651 Node_List* pair = new Node_List();
652 pair->push(t1);
653 pair->push(t2);
654 _packset.append(pair);
655 set_alignment(t1, t2, align);
656 changed = true;
657 }
658 }
659 }
660 return changed;
661 }
662
663 //------------------------------follow_def_uses---------------------------
664 // Extend the packset by visiting uses of nodes in pack p
665 bool SuperWord::follow_def_uses(Node_List* p) {
666 bool changed = false;
667 Node* s1 = p->at(0);
668 Node* s2 = p->at(1);
669 assert(p->size() == 2, "just checking");
670 assert(s1->req() == s2->req(), "just checking");
671 assert(alignment(s1) + data_size(s1) == alignment(s2), "just checking");
672
673 if (s1->is_Store()) return false;
674
675 int align = alignment(s1);
676 int savings = -1;
677 Node* u1 = NULL;
678 Node* u2 = NULL;
679 for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
680 Node* t1 = s1->fast_out(i);
681 if (!in_bb(t1)) continue;
682 for (DUIterator_Fast jmax, j = s2->fast_outs(jmax); j < jmax; j++) {
683 Node* t2 = s2->fast_out(j);
684 if (!in_bb(t2)) continue;
685 if (!opnd_positions_match(s1, t1, s2, t2))
686 continue;
687 if (stmts_can_pack(t1, t2, align)) {
688 int my_savings = est_savings(t1, t2);
689 if (my_savings > savings) {
690 savings = my_savings;
691 u1 = t1;
692 u2 = t2;
693 }
694 }
695 }
696 }
697 if (savings >= 0) {
698 Node_List* pair = new Node_List();
699 pair->push(u1);
700 pair->push(u2);
701 _packset.append(pair);
702 set_alignment(u1, u2, align);
703 changed = true;
704 }
705 return changed;
706 }
707
708 //---------------------------opnd_positions_match-------------------------
709 // Is the use of d1 in u1 at the same operand position as d2 in u2?
710 bool SuperWord::opnd_positions_match(Node* d1, Node* u1, Node* d2, Node* u2) {
711 uint ct = u1->req();
712 if (ct != u2->req()) return false;
713 uint i1 = 0;
714 uint i2 = 0;
715 do {
716 for (i1++; i1 < ct; i1++) if (u1->in(i1) == d1) break;
717 for (i2++; i2 < ct; i2++) if (u2->in(i2) == d2) break;
718 if (i1 != i2) {
719 return false;
720 }
721 } while (i1 < ct);
722 return true;
723 }
724
725 //------------------------------est_savings---------------------------
726 // Estimate the savings from executing s1 and s2 as a pack
727 int SuperWord::est_savings(Node* s1, Node* s2) {
728 int save = 2 - 1; // 2 operations per instruction in packed form
729
730 // inputs
731 for (uint i = 1; i < s1->req(); i++) {
732 Node* x1 = s1->in(i);
733 Node* x2 = s2->in(i);
734 if (x1 != x2) {
735 if (are_adjacent_refs(x1, x2)) {
736 save += adjacent_profit(x1, x2);
737 } else if (!in_packset(x1, x2)) {
738 save -= pack_cost(2);
739 } else {
740 save += unpack_cost(2);
741 }
742 }
743 }
744
745 // uses of result
746 uint ct = 0;
747 for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
748 Node* s1_use = s1->fast_out(i);
749 for (int j = 0; j < _packset.length(); j++) {
750 Node_List* p = _packset.at(j);
751 if (p->at(0) == s1_use) {
752 for (DUIterator_Fast kmax, k = s2->fast_outs(kmax); k < kmax; k++) {
753 Node* s2_use = s2->fast_out(k);
754 if (p->at(p->size()-1) == s2_use) {
755 ct++;
756 if (are_adjacent_refs(s1_use, s2_use)) {
757 save += adjacent_profit(s1_use, s2_use);
758 }
759 }
760 }
761 }
762 }
763 }
764
765 if (ct < s1->outcnt()) save += unpack_cost(1);
766 if (ct < s2->outcnt()) save += unpack_cost(1);
767
768 return save;
769 }
770
771 //------------------------------costs---------------------------
772 int SuperWord::adjacent_profit(Node* s1, Node* s2) { return 2; }
773 int SuperWord::pack_cost(int ct) { return ct; }
774 int SuperWord::unpack_cost(int ct) { return ct; }
775
776 //------------------------------combine_packs---------------------------
777 // Combine packs A and B with A.last == B.first into A.first..,A.last,B.second,..B.last
778 void SuperWord::combine_packs() {
779 bool changed;
780 do {
781 changed = false;
782 for (int i = 0; i < _packset.length(); i++) {
783 Node_List* p1 = _packset.at(i);
784 if (p1 == NULL) continue;
785 for (int j = 0; j < _packset.length(); j++) {
786 Node_List* p2 = _packset.at(j);
787 if (p2 == NULL) continue;
788 if (p1->at(p1->size()-1) == p2->at(0)) {
789 for (uint k = 1; k < p2->size(); k++) {
790 p1->push(p2->at(k));
791 }
792 _packset.at_put(j, NULL);
793 changed = true;
794 }
795 }
796 }
797 } while (changed);
798
799 for (int i = _packset.length() - 1; i >= 0; i--) {
800 Node_List* p1 = _packset.at(i);
801 if (p1 == NULL) {
802 _packset.remove_at(i);
803 }
804 }
805
806 #ifndef PRODUCT
807 if (TraceSuperWord) {
808 tty->print_cr("\nAfter combine_packs");
809 print_packset();
810 }
811 #endif
812 }
813
814 //-----------------------------construct_my_pack_map--------------------------
815 // Construct the map from nodes to packs. Only valid after the
816 // point where a node is only in one pack (after combine_packs).
817 void SuperWord::construct_my_pack_map() {
818 Node_List* rslt = NULL;
819 for (int i = 0; i < _packset.length(); i++) {
820 Node_List* p = _packset.at(i);
821 for (uint j = 0; j < p->size(); j++) {
822 Node* s = p->at(j);
823 assert(my_pack(s) == NULL, "only in one pack");
824 set_my_pack(s, p);
825 }
826 }
827 }
828
829 //------------------------------filter_packs---------------------------
830 // Remove packs that are not implemented or not profitable.
831 void SuperWord::filter_packs() {
832
833 // Remove packs that are not implemented
834 for (int i = _packset.length() - 1; i >= 0; i--) {
835 Node_List* pk = _packset.at(i);
836 bool impl = implemented(pk);
837 if (!impl) {
838 #ifndef PRODUCT
839 if (TraceSuperWord && Verbose) {
840 tty->print_cr("Unimplemented");
841 pk->at(0)->dump();
842 }
843 #endif
844 remove_pack_at(i);
845 }
846 }
847
848 // Remove packs that are not profitable
849 bool changed;
850 do {
851 changed = false;
852 for (int i = _packset.length() - 1; i >= 0; i--) {
853 Node_List* pk = _packset.at(i);
854 bool prof = profitable(pk);
855 if (!prof) {
856 #ifndef PRODUCT
857 if (TraceSuperWord && Verbose) {
858 tty->print_cr("Unprofitable");
859 pk->at(0)->dump();
860 }
861 #endif
862 remove_pack_at(i);
863 changed = true;
864 }
865 }
866 } while (changed);
867
868 #ifndef PRODUCT
869 if (TraceSuperWord) {
870 tty->print_cr("\nAfter filter_packs");
871 print_packset();
872 tty->cr();
873 }
874 #endif
875 }
876
877 //------------------------------implemented---------------------------
878 // Can code be generated for pack p?
879 bool SuperWord::implemented(Node_List* p) {
880 Node* p0 = p->at(0);
881 int vopc = VectorNode::opcode(p0->Opcode(), p->size(), velt_type(p0));
882 return vopc > 0 && Matcher::has_match_rule(vopc);
883 }
884
885 //------------------------------profitable---------------------------
886 // For pack p, are all operands and all uses (with in the block) vector?
887 bool SuperWord::profitable(Node_List* p) {
888 Node* p0 = p->at(0);
889 uint start, end;
890 vector_opd_range(p0, &start, &end);
891
892 // Return false if some input is not vector and inside block
893 for (uint i = start; i < end; i++) {
894 if (!is_vector_use(p0, i)) {
895 // For now, return false if not scalar promotion case (inputs are the same.)
896 // Later, implement PackNode and allow differing, non-vector inputs
897 // (maybe just the ones from outside the block.)
898 Node* p0_def = p0->in(i);
899 for (uint j = 1; j < p->size(); j++) {
900 Node* use = p->at(j);
901 Node* def = use->in(i);
902 if (p0_def != def)
903 return false;
904 }
905 }
906 }
907 if (!p0->is_Store()) {
908 // For now, return false if not all uses are vector.
909 // Later, implement ExtractNode and allow non-vector uses (maybe
910 // just the ones outside the block.)
911 for (uint i = 0; i < p->size(); i++) {
912 Node* def = p->at(i);
913 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
914 Node* use = def->fast_out(j);
915 for (uint k = 0; k < use->req(); k++) {
916 Node* n = use->in(k);
917 if (def == n) {
918 if (!is_vector_use(use, k)) {
919 return false;
920 }
921 }
922 }
923 }
924 }
925 }
926 return true;
927 }
928
929 //------------------------------schedule---------------------------
930 // Adjust the memory graph for the packed operations
931 void SuperWord::schedule() {
932
933 // Co-locate in the memory graph the members of each memory pack
934 for (int i = 0; i < _packset.length(); i++) {
935 co_locate_pack(_packset.at(i));
936 }
937 }
938
939 //-------------------------------remove_and_insert-------------------
940 //remove "current" from its current position in the memory graph and insert
941 //it after the appropriate insertion point (lip or uip)
942 void SuperWord::remove_and_insert(MemNode *current, MemNode *prev, MemNode *lip,
943 Node *uip, Unique_Node_List &sched_before) {
944 Node* my_mem = current->in(MemNode::Memory);
945 _igvn.hash_delete(current);
946 _igvn.hash_delete(my_mem);
947
948 //remove current_store from its current position in the memmory graph
949 for (DUIterator i = current->outs(); current->has_out(i); i++) {
950 Node* use = current->out(i);
951 if (use->is_Mem()) {
952 assert(use->in(MemNode::Memory) == current, "must be");
953 _igvn.hash_delete(use);
954 if (use == prev) { // connect prev to my_mem
955 use->set_req(MemNode::Memory, my_mem);
956 } else if (sched_before.member(use)) {
957 _igvn.hash_delete(uip);
958 use->set_req(MemNode::Memory, uip);
959 } else {
960 _igvn.hash_delete(lip);
961 use->set_req(MemNode::Memory, lip);
962 }
963 _igvn._worklist.push(use);
964 --i; //deleted this edge; rescan position
965 }
966 }
967
968 bool sched_up = sched_before.member(current);
969 Node *insert_pt = sched_up ? uip : lip;
970 _igvn.hash_delete(insert_pt);
971
972 // all uses of insert_pt's memory state should use current's instead
973 for (DUIterator i = insert_pt->outs(); insert_pt->has_out(i); i++) {
974 Node* use = insert_pt->out(i);
975 if (use->is_Mem()) {
976 assert(use->in(MemNode::Memory) == insert_pt, "must be");
977 _igvn.hash_delete(use);
978 use->set_req(MemNode::Memory, current);
979 _igvn._worklist.push(use);
980 --i; //deleted this edge; rescan position
981 } else if (!sched_up && use->is_Phi() && use->bottom_type() == Type::MEMORY) {
982 uint pos; //lip (lower insert point) must be the last one in the memory slice
983 _igvn.hash_delete(use);
984 for (pos=1; pos < use->req(); pos++) {
985 if (use->in(pos) == insert_pt) break;
986 }
987 use->set_req(pos, current);
988 _igvn._worklist.push(use);
989 --i;
990 }
991 }
992
993 //connect current to insert_pt
994 current->set_req(MemNode::Memory, insert_pt);
995 _igvn._worklist.push(current);
996 }
997
998 //------------------------------co_locate_pack----------------------------------
999 // To schedule a store pack, we need to move any sandwiched memory ops either before
1000 // or after the pack, based upon dependence information:
1001 // (1) If any store in the pack depends on the sandwiched memory op, the
1002 // sandwiched memory op must be scheduled BEFORE the pack;
1003 // (2) If a sandwiched memory op depends on any store in the pack, the
1004 // sandwiched memory op must be scheduled AFTER the pack;
1005 // (3) If a sandwiched memory op (say, memA) depends on another sandwiched
1006 // memory op (say memB), memB must be scheduled before memA. So, if memA is
1007 // scheduled before the pack, memB must also be scheduled before the pack;
1008 // (4) If there is no dependence restriction for a sandwiched memory op, we simply
1009 // schedule this store AFTER the pack
1010 // (5) We know there is no dependence cycle, so there in no other case;
1011 // (6) Finally, all memory ops in another single pack should be moved in the same direction.
1012 //
1013 // To schedule a load pack, we use the memory state of either the first or the last load in
1014 // the pack, based on the dependence constraint.
1015 void SuperWord::co_locate_pack(Node_List* pk) {
1016 if (pk->at(0)->is_Store()) {
1017 MemNode* first = executed_first(pk)->as_Mem();
1018 MemNode* last = executed_last(pk)->as_Mem();
1019 Unique_Node_List schedule_before_pack;
1020 Unique_Node_List memops;
1021
1022 MemNode* current = last->in(MemNode::Memory)->as_Mem();
1023 MemNode* previous = last;
1024 while (true) {
1025 assert(in_bb(current), "stay in block");
1026 memops.push(previous);
1027 for (DUIterator i = current->outs(); current->has_out(i); i++) {
1028 Node* use = current->out(i);
1029 if (use->is_Mem() && use != previous)
1030 memops.push(use);
1031 }
1032 if(current == first) break;
1033 previous = current;
1034 current = current->in(MemNode::Memory)->as_Mem();
1035 }
1036
1037 // determine which memory operations should be scheduled before the pack
1038 for (uint i = 1; i < memops.size(); i++) {
1039 Node *s1 = memops.at(i);
1040 if (!in_pack(s1, pk) && !schedule_before_pack.member(s1)) {
1041 for (uint j = 0; j< i; j++) {
1042 Node *s2 = memops.at(j);
1043 if (!independent(s1, s2)) {
1044 if (in_pack(s2, pk) || schedule_before_pack.member(s2)) {
1045 schedule_before_pack.push(s1); //s1 must be scheduled before
1046 Node_List* mem_pk = my_pack(s1);
1047 if (mem_pk != NULL) {
1048 for (uint ii = 0; ii < mem_pk->size(); ii++) {
1049 Node* s = mem_pk->at(ii); // follow partner
1050 if (memops.member(s) && !schedule_before_pack.member(s))
1051 schedule_before_pack.push(s);
1052 }
1053 }
1054 }
1055 }
1056 }
1057 }
1058 }
1059
1060 MemNode* lower_insert_pt = last;
1061 Node* upper_insert_pt = first->in(MemNode::Memory);
1062 previous = last; //previous store in pk
1063 current = last->in(MemNode::Memory)->as_Mem();
1064
1065 //start scheduling from "last" to "first"
1066 while (true) {
1067 assert(in_bb(current), "stay in block");
1068 assert(in_pack(previous, pk), "previous stays in pack");
1069 Node* my_mem = current->in(MemNode::Memory);
1070
1071 if (in_pack(current, pk)) {
1072 // Forward users of my memory state (except "previous) to my input memory state
1073 _igvn.hash_delete(current);
1074 for (DUIterator i = current->outs(); current->has_out(i); i++) {
1075 Node* use = current->out(i);
1076 if (use->is_Mem() && use != previous) {
1077 assert(use->in(MemNode::Memory) == current, "must be");
1078 _igvn.hash_delete(use);
1079 if (schedule_before_pack.member(use)) {
1080 _igvn.hash_delete(upper_insert_pt);
1081 use->set_req(MemNode::Memory, upper_insert_pt);
1082 } else {
1083 _igvn.hash_delete(lower_insert_pt);
1084 use->set_req(MemNode::Memory, lower_insert_pt);
1085 }
1086 _igvn._worklist.push(use);
1087 --i; // deleted this edge; rescan position
1088 }
1089 }
1090 previous = current;
1091 } else { // !in_pack(current, pk) ==> a sandwiched store
1092 remove_and_insert(current, previous, lower_insert_pt, upper_insert_pt, schedule_before_pack);
1093 }
1094
1095 if (current == first) break;
1096 current = my_mem->as_Mem();
1097 } // end while
1098 } else if (pk->at(0)->is_Load()) { //load
1099 // all loads in the pack should have the same memory state. By default,
1100 // we use the memory state of the last load. However, if any load could
1101 // not be moved down due to the dependence constraint, we use the memory
1102 // state of the first load.
1103 Node* last_mem = executed_last(pk)->in(MemNode::Memory);
1104 Node* first_mem = executed_first(pk)->in(MemNode::Memory);
1105 bool schedule_last = true;
1106 for (uint i = 0; i < pk->size(); i++) {
1107 Node* ld = pk->at(i);
1108 for (Node* current = last_mem; current != ld->in(MemNode::Memory);
1109 current=current->in(MemNode::Memory)) {
1110 assert(current != first_mem, "corrupted memory graph");
1111 if(current->is_Mem() && !independent(current, ld)){
1112 schedule_last = false; // a later store depends on this load
1113 break;
1114 }
1115 }
1116 }
1117
1118 Node* mem_input = schedule_last ? last_mem : first_mem;
1119 _igvn.hash_delete(mem_input);
1120 // Give each load the same memory state
1121 for (uint i = 0; i < pk->size(); i++) {
1122 LoadNode* ld = pk->at(i)->as_Load();
1123 _igvn.hash_delete(ld);
1124 ld->set_req(MemNode::Memory, mem_input);
1125 _igvn._worklist.push(ld);
1126 }
1127 }
1128 }
1129
1130 //------------------------------output---------------------------
1131 // Convert packs into vector node operations
1132 void SuperWord::output() {
1133 if (_packset.length() == 0) return;
1134
1135 #ifndef PRODUCT
1136 if (TraceLoopOpts) {
1137 tty->print("SuperWord ");
1138 lpt()->dump_head();
1139 }
1140 #endif
1141
1142 // MUST ENSURE main loop's initial value is properly aligned:
1143 // (iv_initial_value + min_iv_offset) % vector_width_in_bytes() == 0
1144
1145 align_initial_loop_index(align_to_ref());
1146
1147 // Insert extract (unpack) operations for scalar uses
1148 for (int i = 0; i < _packset.length(); i++) {
1149 insert_extracts(_packset.at(i));
1150 }
1151
1152 for (int i = 0; i < _block.length(); i++) {
1153 Node* n = _block.at(i);
1154 Node_List* p = my_pack(n);
1155 if (p && n == executed_last(p)) {
1156 uint vlen = p->size();
1157 Node* vn = NULL;
1158 Node* low_adr = p->at(0);
1159 Node* first = executed_first(p);
1160 if (n->is_Load()) {
1161 int opc = n->Opcode();
1162 Node* ctl = n->in(MemNode::Control);
1163 Node* mem = first->in(MemNode::Memory);
1164 Node* adr = low_adr->in(MemNode::Address);
1165 const TypePtr* atyp = n->adr_type();
1166 vn = VectorLoadNode::make(_phase->C, opc, ctl, mem, adr, atyp, vlen);
1167
1168 } else if (n->is_Store()) {
1169 // Promote value to be stored to vector
1170 VectorNode* val = vector_opd(p, MemNode::ValueIn);
1171
1172 int opc = n->Opcode();
1173 Node* ctl = n->in(MemNode::Control);
1174 Node* mem = first->in(MemNode::Memory);
1175 Node* adr = low_adr->in(MemNode::Address);
1176 const TypePtr* atyp = n->adr_type();
1177 vn = VectorStoreNode::make(_phase->C, opc, ctl, mem, adr, atyp, val, vlen);
1178
1179 } else if (n->req() == 3) {
1180 // Promote operands to vector
1181 Node* in1 = vector_opd(p, 1);
1182 Node* in2 = vector_opd(p, 2);
1183 vn = VectorNode::make(_phase->C, n->Opcode(), in1, in2, vlen, velt_type(n));
1184
1185 } else {
1186 ShouldNotReachHere();
1187 }
1188
1189 _phase->_igvn.register_new_node_with_optimizer(vn);
1190 _phase->set_ctrl(vn, _phase->get_ctrl(p->at(0)));
1191 for (uint j = 0; j < p->size(); j++) {
1192 Node* pm = p->at(j);
1193 _igvn.replace_node(pm, vn);
1194 }
1195 _igvn._worklist.push(vn);
1196 }
1197 }
1198 }
1199
1200 //------------------------------vector_opd---------------------------
1201 // Create a vector operand for the nodes in pack p for operand: in(opd_idx)
1202 VectorNode* SuperWord::vector_opd(Node_List* p, int opd_idx) {
1203 Node* p0 = p->at(0);
1204 uint vlen = p->size();
1205 Node* opd = p0->in(opd_idx);
1206
1207 bool same_opd = true;
1208 for (uint i = 1; i < vlen; i++) {
1209 Node* pi = p->at(i);
1210 Node* in = pi->in(opd_idx);
1211 if (opd != in) {
1212 same_opd = false;
1213 break;
1214 }
1215 }
1216
1217 if (same_opd) {
1218 if (opd->is_Vector()) {
1219 return (VectorNode*)opd; // input is matching vector
1220 }
1221 // Convert scalar input to vector. Use p0's type because it's container
1222 // maybe smaller than the operand's container.
1223 const Type* opd_t = velt_type(!in_bb(opd) ? p0 : opd);
1224 const Type* p0_t = velt_type(p0);
1225 if (p0_t->higher_equal(opd_t)) opd_t = p0_t;
1226 VectorNode* vn = VectorNode::scalar2vector(_phase->C, opd, vlen, opd_t);
1227
1228 _phase->_igvn.register_new_node_with_optimizer(vn);
1229 _phase->set_ctrl(vn, _phase->get_ctrl(opd));
1230 return vn;
1231 }
1232
1233 // Insert pack operation
1234 const Type* opd_t = velt_type(!in_bb(opd) ? p0 : opd);
1235 PackNode* pk = PackNode::make(_phase->C, opd, opd_t);
1236
1237 for (uint i = 1; i < vlen; i++) {
1238 Node* pi = p->at(i);
1239 Node* in = pi->in(opd_idx);
1240 assert(my_pack(in) == NULL, "Should already have been unpacked");
1241 assert(opd_t == velt_type(!in_bb(in) ? pi : in), "all same type");
1242 pk->add_opd(in);
1243 }
1244 _phase->_igvn.register_new_node_with_optimizer(pk);
1245 _phase->set_ctrl(pk, _phase->get_ctrl(opd));
1246 return pk;
1247 }
1248
1249 //------------------------------insert_extracts---------------------------
1250 // If a use of pack p is not a vector use, then replace the
1251 // use with an extract operation.
1252 void SuperWord::insert_extracts(Node_List* p) {
1253 if (p->at(0)->is_Store()) return;
1254 assert(_n_idx_list.is_empty(), "empty (node,index) list");
1255
1256 // Inspect each use of each pack member. For each use that is
1257 // not a vector use, replace the use with an extract operation.
1258
1259 for (uint i = 0; i < p->size(); i++) {
1260 Node* def = p->at(i);
1261 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
1262 Node* use = def->fast_out(j);
1263 for (uint k = 0; k < use->req(); k++) {
1264 Node* n = use->in(k);
1265 if (def == n) {
1266 if (!is_vector_use(use, k)) {
1267 _n_idx_list.push(use, k);
1268 }
1269 }
1270 }
1271 }
1272 }
1273
1274 while (_n_idx_list.is_nonempty()) {
1275 Node* use = _n_idx_list.node();
1276 int idx = _n_idx_list.index();
1277 _n_idx_list.pop();
1278 Node* def = use->in(idx);
1279
1280 // Insert extract operation
1281 _igvn.hash_delete(def);
1282 _igvn.hash_delete(use);
1283 int def_pos = alignment(def) / data_size(def);
1284 const Type* def_t = velt_type(def);
1285
1286 Node* ex = ExtractNode::make(_phase->C, def, def_pos, def_t);
1287 _phase->_igvn.register_new_node_with_optimizer(ex);
1288 _phase->set_ctrl(ex, _phase->get_ctrl(def));
1289 use->set_req(idx, ex);
1290 _igvn._worklist.push(def);
1291 _igvn._worklist.push(use);
1292
1293 bb_insert_after(ex, bb_idx(def));
1294 set_velt_type(ex, def_t);
1295 }
1296 }
1297
1298 //------------------------------is_vector_use---------------------------
1299 // Is use->in(u_idx) a vector use?
1300 bool SuperWord::is_vector_use(Node* use, int u_idx) {
1301 Node_List* u_pk = my_pack(use);
1302 if (u_pk == NULL) return false;
1303 Node* def = use->in(u_idx);
1304 Node_List* d_pk = my_pack(def);
1305 if (d_pk == NULL) {
1306 // check for scalar promotion
1307 Node* n = u_pk->at(0)->in(u_idx);
1308 for (uint i = 1; i < u_pk->size(); i++) {
1309 if (u_pk->at(i)->in(u_idx) != n) return false;
1310 }
1311 return true;
1312 }
1313 if (u_pk->size() != d_pk->size())
1314 return false;
1315 for (uint i = 0; i < u_pk->size(); i++) {
1316 Node* ui = u_pk->at(i);
1317 Node* di = d_pk->at(i);
1318 if (ui->in(u_idx) != di || alignment(ui) != alignment(di))
1319 return false;
1320 }
1321 return true;
1322 }
1323
1324 //------------------------------construct_bb---------------------------
1325 // Construct reverse postorder list of block members
1326 void SuperWord::construct_bb() {
1327 Node* entry = bb();
1328
1329 assert(_stk.length() == 0, "stk is empty");
1330 assert(_block.length() == 0, "block is empty");
1331 assert(_data_entry.length() == 0, "data_entry is empty");
1332 assert(_mem_slice_head.length() == 0, "mem_slice_head is empty");
1333 assert(_mem_slice_tail.length() == 0, "mem_slice_tail is empty");
1334
1335 // Find non-control nodes with no inputs from within block,
1336 // create a temporary map from node _idx to bb_idx for use
1337 // by the visited and post_visited sets,
1338 // and count number of nodes in block.
1339 int bb_ct = 0;
1340 for (uint i = 0; i < lpt()->_body.size(); i++ ) {
1341 Node *n = lpt()->_body.at(i);
1342 set_bb_idx(n, i); // Create a temporary map
1343 if (in_bb(n)) {
1344 bb_ct++;
1345 if (!n->is_CFG()) {
1346 bool found = false;
1347 for (uint j = 0; j < n->req(); j++) {
1348 Node* def = n->in(j);
1349 if (def && in_bb(def)) {
1350 found = true;
1351 break;
1352 }
1353 }
1354 if (!found) {
1355 assert(n != entry, "can't be entry");
1356 _data_entry.push(n);
1357 }
1358 }
1359 }
1360 }
1361
1362 // Find memory slices (head and tail)
1363 for (DUIterator_Fast imax, i = lp()->fast_outs(imax); i < imax; i++) {
1364 Node *n = lp()->fast_out(i);
1365 if (in_bb(n) && (n->is_Phi() && n->bottom_type() == Type::MEMORY)) {
1366 Node* n_tail = n->in(LoopNode::LoopBackControl);
1367 if (n_tail != n->in(LoopNode::EntryControl)) {
1368 _mem_slice_head.push(n);
1369 _mem_slice_tail.push(n_tail);
1370 }
1371 }
1372 }
1373
1374 // Create an RPO list of nodes in block
1375
1376 visited_clear();
1377 post_visited_clear();
1378
1379 // Push all non-control nodes with no inputs from within block, then control entry
1380 for (int j = 0; j < _data_entry.length(); j++) {
1381 Node* n = _data_entry.at(j);
1382 visited_set(n);
1383 _stk.push(n);
1384 }
1385 visited_set(entry);
1386 _stk.push(entry);
1387
1388 // Do a depth first walk over out edges
1389 int rpo_idx = bb_ct - 1;
1390 int size;
1391 while ((size = _stk.length()) > 0) {
1392 Node* n = _stk.top(); // Leave node on stack
1393 if (!visited_test_set(n)) {
1394 // forward arc in graph
1395 } else if (!post_visited_test(n)) {
1396 // cross or back arc
1397 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1398 Node *use = n->fast_out(i);
1399 if (in_bb(use) && !visited_test(use) &&
1400 // Don't go around backedge
1401 (!use->is_Phi() || n == entry)) {
1402 _stk.push(use);
1403 }
1404 }
1405 if (_stk.length() == size) {
1406 // There were no additional uses, post visit node now
1407 _stk.pop(); // Remove node from stack
1408 assert(rpo_idx >= 0, "");
1409 _block.at_put_grow(rpo_idx, n);
1410 rpo_idx--;
1411 post_visited_set(n);
1412 assert(rpo_idx >= 0 || _stk.is_empty(), "");
1413 }
1414 } else {
1415 _stk.pop(); // Remove post-visited node from stack
1416 }
1417 }
1418
1419 // Create real map of block indices for nodes
1420 for (int j = 0; j < _block.length(); j++) {
1421 Node* n = _block.at(j);
1422 set_bb_idx(n, j);
1423 }
1424
1425 initialize_bb(); // Ensure extra info is allocated.
1426
1427 #ifndef PRODUCT
1428 if (TraceSuperWord) {
1429 print_bb();
1430 tty->print_cr("\ndata entry nodes: %s", _data_entry.length() > 0 ? "" : "NONE");
1431 for (int m = 0; m < _data_entry.length(); m++) {
1432 tty->print("%3d ", m);
1433 _data_entry.at(m)->dump();
1434 }
1435 tty->print_cr("\nmemory slices: %s", _mem_slice_head.length() > 0 ? "" : "NONE");
1436 for (int m = 0; m < _mem_slice_head.length(); m++) {
1437 tty->print("%3d ", m); _mem_slice_head.at(m)->dump();
1438 tty->print(" "); _mem_slice_tail.at(m)->dump();
1439 }
1440 }
1441 #endif
1442 assert(rpo_idx == -1 && bb_ct == _block.length(), "all block members found");
1443 }
1444
1445 //------------------------------initialize_bb---------------------------
1446 // Initialize per node info
1447 void SuperWord::initialize_bb() {
1448 Node* last = _block.at(_block.length() - 1);
1449 grow_node_info(bb_idx(last));
1450 }
1451
1452 //------------------------------bb_insert_after---------------------------
1453 // Insert n into block after pos
1454 void SuperWord::bb_insert_after(Node* n, int pos) {
1455 int n_pos = pos + 1;
1456 // Make room
1457 for (int i = _block.length() - 1; i >= n_pos; i--) {
1458 _block.at_put_grow(i+1, _block.at(i));
1459 }
1460 for (int j = _node_info.length() - 1; j >= n_pos; j--) {
1461 _node_info.at_put_grow(j+1, _node_info.at(j));
1462 }
1463 // Set value
1464 _block.at_put_grow(n_pos, n);
1465 _node_info.at_put_grow(n_pos, SWNodeInfo::initial);
1466 // Adjust map from node->_idx to _block index
1467 for (int i = n_pos; i < _block.length(); i++) {
1468 set_bb_idx(_block.at(i), i);
1469 }
1470 }
1471
1472 //------------------------------compute_max_depth---------------------------
1473 // Compute max depth for expressions from beginning of block
1474 // Use to prune search paths during test for independence.
1475 void SuperWord::compute_max_depth() {
1476 int ct = 0;
1477 bool again;
1478 do {
1479 again = false;
1480 for (int i = 0; i < _block.length(); i++) {
1481 Node* n = _block.at(i);
1482 if (!n->is_Phi()) {
1483 int d_orig = depth(n);
1484 int d_in = 0;
1485 for (DepPreds preds(n, _dg); !preds.done(); preds.next()) {
1486 Node* pred = preds.current();
1487 if (in_bb(pred)) {
1488 d_in = MAX2(d_in, depth(pred));
1489 }
1490 }
1491 if (d_in + 1 != d_orig) {
1492 set_depth(n, d_in + 1);
1493 again = true;
1494 }
1495 }
1496 }
1497 ct++;
1498 } while (again);
1499 #ifndef PRODUCT
1500 if (TraceSuperWord && Verbose)
1501 tty->print_cr("compute_max_depth iterated: %d times", ct);
1502 #endif
1503 }
1504
1505 //-------------------------compute_vector_element_type-----------------------
1506 // Compute necessary vector element type for expressions
1507 // This propagates backwards a narrower integer type when the
1508 // upper bits of the value are not needed.
1509 // Example: char a,b,c; a = b + c;
1510 // Normally the type of the add is integer, but for packed character
1511 // operations the type of the add needs to be char.
1512 void SuperWord::compute_vector_element_type() {
1513 #ifndef PRODUCT
1514 if (TraceSuperWord && Verbose)
1515 tty->print_cr("\ncompute_velt_type:");
1516 #endif
1517
1518 // Initial type
1519 for (int i = 0; i < _block.length(); i++) {
1520 Node* n = _block.at(i);
1521 const Type* t = n->is_Mem() ? Type::get_const_basic_type(n->as_Mem()->memory_type())
1522 : _igvn.type(n);
1523 const Type* vt = container_type(t);
1524 set_velt_type(n, vt);
1525 }
1526
1527 // Propagate narrowed type backwards through operations
1528 // that don't depend on higher order bits
1529 for (int i = _block.length() - 1; i >= 0; i--) {
1530 Node* n = _block.at(i);
1531 // Only integer types need be examined
1532 if (n->bottom_type()->isa_int()) {
1533 uint start, end;
1534 vector_opd_range(n, &start, &end);
1535 const Type* vt = velt_type(n);
1536
1537 for (uint j = start; j < end; j++) {
1538 Node* in = n->in(j);
1539 // Don't propagate through a type conversion
1540 if (n->bottom_type() != in->bottom_type())
1541 continue;
1542 switch(in->Opcode()) {
1543 case Op_AddI: case Op_AddL:
1544 case Op_SubI: case Op_SubL:
1545 case Op_MulI: case Op_MulL:
1546 case Op_AndI: case Op_AndL:
1547 case Op_OrI: case Op_OrL:
1548 case Op_XorI: case Op_XorL:
1549 case Op_LShiftI: case Op_LShiftL:
1550 case Op_CMoveI: case Op_CMoveL:
1551 if (in_bb(in)) {
1552 bool same_type = true;
1553 for (DUIterator_Fast kmax, k = in->fast_outs(kmax); k < kmax; k++) {
1554 Node *use = in->fast_out(k);
1555 if (!in_bb(use) || velt_type(use) != vt) {
1556 same_type = false;
1557 break;
1558 }
1559 }
1560 if (same_type) {
1561 set_velt_type(in, vt);
1562 }
1563 }
1564 }
1565 }
1566 }
1567 }
1568 #ifndef PRODUCT
1569 if (TraceSuperWord && Verbose) {
1570 for (int i = 0; i < _block.length(); i++) {
1571 Node* n = _block.at(i);
1572 velt_type(n)->dump();
1573 tty->print("\t");
1574 n->dump();
1575 }
1576 }
1577 #endif
1578 }
1579
1580 //------------------------------memory_alignment---------------------------
1581 // Alignment within a vector memory reference
1582 int SuperWord::memory_alignment(MemNode* s, int iv_adjust_in_bytes) {
1583 SWPointer p(s, this);
1584 if (!p.valid()) {
1585 return bottom_align;
1586 }
1587 int offset = p.offset_in_bytes();
1588 offset += iv_adjust_in_bytes;
1589 int off_rem = offset % vector_width_in_bytes();
1590 int off_mod = off_rem >= 0 ? off_rem : off_rem + vector_width_in_bytes();
1591 return off_mod;
1592 }
1593
1594 //---------------------------container_type---------------------------
1595 // Smallest type containing range of values
1596 const Type* SuperWord::container_type(const Type* t) {
1597 const Type* tp = t->make_ptr();
1598 if (tp && tp->isa_aryptr()) {
1599 t = tp->is_aryptr()->elem();
1600 }
1601 if (t->basic_type() == T_INT) {
1602 if (t->higher_equal(TypeInt::BOOL)) return TypeInt::BOOL;
1603 if (t->higher_equal(TypeInt::BYTE)) return TypeInt::BYTE;
1604 if (t->higher_equal(TypeInt::CHAR)) return TypeInt::CHAR;
1605 if (t->higher_equal(TypeInt::SHORT)) return TypeInt::SHORT;
1606 return TypeInt::INT;
1607 }
1608 return t;
1609 }
1610
1611 //-------------------------vector_opd_range-----------------------
1612 // (Start, end] half-open range defining which operands are vector
1613 void SuperWord::vector_opd_range(Node* n, uint* start, uint* end) {
1614 switch (n->Opcode()) {
1615 case Op_LoadB: case Op_LoadUS:
1616 case Op_LoadI: case Op_LoadL:
1617 case Op_LoadF: case Op_LoadD:
1618 case Op_LoadP:
1619 *start = 0;
1620 *end = 0;
1621 return;
1622 case Op_StoreB: case Op_StoreC:
1623 case Op_StoreI: case Op_StoreL:
1624 case Op_StoreF: case Op_StoreD:
1625 case Op_StoreP:
1626 *start = MemNode::ValueIn;
1627 *end = *start + 1;
1628 return;
1629 case Op_LShiftI: case Op_LShiftL:
1630 *start = 1;
1631 *end = 2;
1632 return;
1633 case Op_CMoveI: case Op_CMoveL: case Op_CMoveF: case Op_CMoveD:
1634 *start = 2;
1635 *end = n->req();
1636 return;
1637 }
1638 *start = 1;
1639 *end = n->req(); // default is all operands
1640 }
1641
1642 //------------------------------in_packset---------------------------
1643 // Are s1 and s2 in a pack pair and ordered as s1,s2?
1644 bool SuperWord::in_packset(Node* s1, Node* s2) {
1645 for (int i = 0; i < _packset.length(); i++) {
1646 Node_List* p = _packset.at(i);
1647 assert(p->size() == 2, "must be");
1648 if (p->at(0) == s1 && p->at(p->size()-1) == s2) {
1649 return true;
1650 }
1651 }
1652 return false;
1653 }
1654
1655 //------------------------------in_pack---------------------------
1656 // Is s in pack p?
1657 Node_List* SuperWord::in_pack(Node* s, Node_List* p) {
1658 for (uint i = 0; i < p->size(); i++) {
1659 if (p->at(i) == s) {
1660 return p;
1661 }
1662 }
1663 return NULL;
1664 }
1665
1666 //------------------------------remove_pack_at---------------------------
1667 // Remove the pack at position pos in the packset
1668 void SuperWord::remove_pack_at(int pos) {
1669 Node_List* p = _packset.at(pos);
1670 for (uint i = 0; i < p->size(); i++) {
1671 Node* s = p->at(i);
1672 set_my_pack(s, NULL);
1673 }
1674 _packset.remove_at(pos);
1675 }
1676
1677 //------------------------------executed_first---------------------------
1678 // Return the node executed first in pack p. Uses the RPO block list
1679 // to determine order.
1680 Node* SuperWord::executed_first(Node_List* p) {
1681 Node* n = p->at(0);
1682 int n_rpo = bb_idx(n);
1683 for (uint i = 1; i < p->size(); i++) {
1684 Node* s = p->at(i);
1685 int s_rpo = bb_idx(s);
1686 if (s_rpo < n_rpo) {
1687 n = s;
1688 n_rpo = s_rpo;
1689 }
1690 }
1691 return n;
1692 }
1693
1694 //------------------------------executed_last---------------------------
1695 // Return the node executed last in pack p.
1696 Node* SuperWord::executed_last(Node_List* p) {
1697 Node* n = p->at(0);
1698 int n_rpo = bb_idx(n);
1699 for (uint i = 1; i < p->size(); i++) {
1700 Node* s = p->at(i);
1701 int s_rpo = bb_idx(s);
1702 if (s_rpo > n_rpo) {
1703 n = s;
1704 n_rpo = s_rpo;
1705 }
1706 }
1707 return n;
1708 }
1709
1710 //----------------------------align_initial_loop_index---------------------------
1711 // Adjust pre-loop limit so that in main loop, a load/store reference
1712 // to align_to_ref will be a position zero in the vector.
1713 // (iv + k) mod vector_align == 0
1714 void SuperWord::align_initial_loop_index(MemNode* align_to_ref) {
1715 CountedLoopNode *main_head = lp()->as_CountedLoop();
1716 assert(main_head->is_main_loop(), "");
1717 CountedLoopEndNode* pre_end = get_pre_loop_end(main_head);
1718 assert(pre_end != NULL, "");
1719 Node *pre_opaq1 = pre_end->limit();
1720 assert(pre_opaq1->Opcode() == Op_Opaque1, "");
1721 Opaque1Node *pre_opaq = (Opaque1Node*)pre_opaq1;
1722 Node *lim0 = pre_opaq->in(1);
1723
1724 // Where we put new limit calculations
1725 Node *pre_ctrl = pre_end->loopnode()->in(LoopNode::EntryControl);
1726
1727 // Ensure the original loop limit is available from the
1728 // pre-loop Opaque1 node.
1729 Node *orig_limit = pre_opaq->original_loop_limit();
1730 assert(orig_limit != NULL && _igvn.type(orig_limit) != Type::TOP, "");
1731
1732 SWPointer align_to_ref_p(align_to_ref, this);
1733
1734 // Given:
1735 // lim0 == original pre loop limit
1736 // V == v_align (power of 2)
1737 // invar == extra invariant piece of the address expression
1738 // e == k [ +/- invar ]
1739 //
1740 // When reassociating expressions involving '%' the basic rules are:
1741 // (a - b) % k == 0 => a % k == b % k
1742 // and:
1743 // (a + b) % k == 0 => a % k == (k - b) % k
1744 //
1745 // For stride > 0 && scale > 0,
1746 // Derive the new pre-loop limit "lim" such that the two constraints:
1747 // (1) lim = lim0 + N (where N is some positive integer < V)
1748 // (2) (e + lim) % V == 0
1749 // are true.
1750 //
1751 // Substituting (1) into (2),
1752 // (e + lim0 + N) % V == 0
1753 // solve for N:
1754 // N = (V - (e + lim0)) % V
1755 // substitute back into (1), so that new limit
1756 // lim = lim0 + (V - (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 = (e - lim0) % V
1765 // lim = lim0 + (e - lim0) % V
1766 //
1767 // For stride < 0 && scale > 0
1768 // Constraints:
1769 // lim = lim0 - N
1770 // (e + lim) % V == 0
1771 // Solving for lim:
1772 // (e + lim0 - N) % V == 0
1773 // N = (e + lim0) % V
1774 // lim = lim0 - (e + lim0) % V
1775 //
1776 // For stride < 0 && scale < 0
1777 // Constraints:
1778 // lim = lim0 - N
1779 // (e - lim) % V == 0
1780 // Solving for lim:
1781 // (e - lim0 + N) % V == 0
1782 // N = (V - (e - lim0)) % V
1783 // lim = lim0 - (V - (e - lim0)) % V
1784
1785 int stride = iv_stride();
1786 int scale = align_to_ref_p.scale_in_bytes();
1787 int elt_size = align_to_ref_p.memory_size();
1788 int v_align = vector_width_in_bytes() / elt_size;
1789 int k = align_to_ref_p.offset_in_bytes() / elt_size;
1790
1791 Node *kn = _igvn.intcon(k);
1792
1793 Node *e = kn;
1794 if (align_to_ref_p.invar() != NULL) {
1795 // incorporate any extra invariant piece producing k +/- invar >>> log2(elt)
1796 Node* log2_elt = _igvn.intcon(exact_log2(elt_size));
1797 Node* aref = new (_phase->C, 3) URShiftINode(align_to_ref_p.invar(), log2_elt);
1798 _phase->_igvn.register_new_node_with_optimizer(aref);
1799 _phase->set_ctrl(aref, pre_ctrl);
1800 if (align_to_ref_p.negate_invar()) {
1801 e = new (_phase->C, 3) SubINode(e, aref);
1802 } else {
1803 e = new (_phase->C, 3) AddINode(e, aref);
1804 }
1805 _phase->_igvn.register_new_node_with_optimizer(e);
1806 _phase->set_ctrl(e, pre_ctrl);
1807 }
1808
1809 // compute e +/- lim0
1810 if (scale < 0) {
1811 e = new (_phase->C, 3) SubINode(e, lim0);
1812 } else {
1813 e = new (_phase->C, 3) AddINode(e, lim0);
1814 }
1815 _phase->_igvn.register_new_node_with_optimizer(e);
1816 _phase->set_ctrl(e, pre_ctrl);
1817
1818 if (stride * scale > 0) {
1819 // compute V - (e +/- lim0)
1820 Node* va = _igvn.intcon(v_align);
1821 e = new (_phase->C, 3) SubINode(va, e);
1822 _phase->_igvn.register_new_node_with_optimizer(e);
1823 _phase->set_ctrl(e, pre_ctrl);
1824 }
1825 // compute N = (exp) % V
1826 Node* va_msk = _igvn.intcon(v_align - 1);
1827 Node* N = new (_phase->C, 3) AndINode(e, va_msk);
1828 _phase->_igvn.register_new_node_with_optimizer(N);
1829 _phase->set_ctrl(N, pre_ctrl);
1830
1831 // substitute back into (1), so that new limit
1832 // lim = lim0 + N
1833 Node* lim;
1834 if (stride < 0) {
1835 lim = new (_phase->C, 3) SubINode(lim0, N);
1836 } else {
1837 lim = new (_phase->C, 3) AddINode(lim0, N);
1838 }
1839 _phase->_igvn.register_new_node_with_optimizer(lim);
1840 _phase->set_ctrl(lim, pre_ctrl);
1841 Node* constrained =
1842 (stride > 0) ? (Node*) new (_phase->C,3) MinINode(lim, orig_limit)
1843 : (Node*) new (_phase->C,3) MaxINode(lim, orig_limit);
1844 _phase->_igvn.register_new_node_with_optimizer(constrained);
1845 _phase->set_ctrl(constrained, pre_ctrl);
1846 _igvn.hash_delete(pre_opaq);
1847 pre_opaq->set_req(1, constrained);
1848 }
1849
1850 //----------------------------get_pre_loop_end---------------------------
1851 // Find pre loop end from main loop. Returns null if none.
1852 CountedLoopEndNode* SuperWord::get_pre_loop_end(CountedLoopNode *cl) {
1853 Node *ctrl = cl->in(LoopNode::EntryControl);
1854 if (!ctrl->is_IfTrue() && !ctrl->is_IfFalse()) return NULL;
1855 Node *iffm = ctrl->in(0);
1856 if (!iffm->is_If()) return NULL;
1857 Node *p_f = iffm->in(0);
1858 if (!p_f->is_IfFalse()) return NULL;
1859 if (!p_f->in(0)->is_CountedLoopEnd()) return NULL;
1860 CountedLoopEndNode *pre_end = p_f->in(0)->as_CountedLoopEnd();
1861 if (!pre_end->loopnode()->is_pre_loop()) return NULL;
1862 return pre_end;
1863 }
1864
1865
1866 //------------------------------init---------------------------
1867 void SuperWord::init() {
1868 _dg.init();
1869 _packset.clear();
1870 _disjoint_ptrs.clear();
1871 _block.clear();
1872 _data_entry.clear();
1873 _mem_slice_head.clear();
1874 _mem_slice_tail.clear();
1875 _node_info.clear();
1876 _align_to_ref = NULL;
1877 _lpt = NULL;
1878 _lp = NULL;
1879 _bb = NULL;
1880 _iv = NULL;
1881 }
1882
1883 //------------------------------print_packset---------------------------
1884 void SuperWord::print_packset() {
1885 #ifndef PRODUCT
1886 tty->print_cr("packset");
1887 for (int i = 0; i < _packset.length(); i++) {
1888 tty->print_cr("Pack: %d", i);
1889 Node_List* p = _packset.at(i);
1890 print_pack(p);
1891 }
1892 #endif
1893 }
1894
1895 //------------------------------print_pack---------------------------
1896 void SuperWord::print_pack(Node_List* p) {
1897 for (uint i = 0; i < p->size(); i++) {
1898 print_stmt(p->at(i));
1899 }
1900 }
1901
1902 //------------------------------print_bb---------------------------
1903 void SuperWord::print_bb() {
1904 #ifndef PRODUCT
1905 tty->print_cr("\nBlock");
1906 for (int i = 0; i < _block.length(); i++) {
1907 Node* n = _block.at(i);
1908 tty->print("%d ", i);
1909 if (n) {
1910 n->dump();
1911 }
1912 }
1913 #endif
1914 }
1915
1916 //------------------------------print_stmt---------------------------
1917 void SuperWord::print_stmt(Node* s) {
1918 #ifndef PRODUCT
1919 tty->print(" align: %d \t", alignment(s));
1920 s->dump();
1921 #endif
1922 }
1923
1924 //------------------------------blank---------------------------
1925 char* SuperWord::blank(uint depth) {
1926 static char blanks[101];
1927 assert(depth < 101, "too deep");
1928 for (uint i = 0; i < depth; i++) blanks[i] = ' ';
1929 blanks[depth] = '\0';
1930 return blanks;
1931 }
1932
1933
1934 //==============================SWPointer===========================
1935
1936 //----------------------------SWPointer------------------------
1937 SWPointer::SWPointer(MemNode* mem, SuperWord* slp) :
1938 _mem(mem), _slp(slp), _base(NULL), _adr(NULL),
1939 _scale(0), _offset(0), _invar(NULL), _negate_invar(false) {
1940
1941 Node* adr = mem->in(MemNode::Address);
1942 if (!adr->is_AddP()) {
1943 assert(!valid(), "too complex");
1944 return;
1945 }
1946 // Match AddP(base, AddP(ptr, k*iv [+ invariant]), constant)
1947 Node* base = adr->in(AddPNode::Base);
1948 //unsafe reference could not be aligned appropriately without runtime checking
1949 if (base == NULL || base->bottom_type() == Type::TOP) {
1950 assert(!valid(), "unsafe access");
1951 return;
1952 }
1953 for (int i = 0; i < 3; i++) {
1954 if (!scaled_iv_plus_offset(adr->in(AddPNode::Offset))) {
1955 assert(!valid(), "too complex");
1956 return;
1957 }
1958 adr = adr->in(AddPNode::Address);
1959 if (base == adr || !adr->is_AddP()) {
1960 break; // stop looking at addp's
1961 }
1962 }
1963 _base = base;
1964 _adr = adr;
1965 assert(valid(), "Usable");
1966 }
1967
1968 // Following is used to create a temporary object during
1969 // the pattern match of an address expression.
1970 SWPointer::SWPointer(SWPointer* p) :
1971 _mem(p->_mem), _slp(p->_slp), _base(NULL), _adr(NULL),
1972 _scale(0), _offset(0), _invar(NULL), _negate_invar(false) {}
1973
1974 //------------------------scaled_iv_plus_offset--------------------
1975 // Match: k*iv + offset
1976 // where: k is a constant that maybe zero, and
1977 // offset is (k2 [+/- invariant]) where k2 maybe zero and invariant is optional
1978 bool SWPointer::scaled_iv_plus_offset(Node* n) {
1979 if (scaled_iv(n)) {
1980 return true;
1981 }
1982 if (offset_plus_k(n)) {
1983 return true;
1984 }
1985 int opc = n->Opcode();
1986 if (opc == Op_AddI) {
1987 if (scaled_iv(n->in(1)) && offset_plus_k(n->in(2))) {
1988 return true;
1989 }
1990 if (scaled_iv(n->in(2)) && offset_plus_k(n->in(1))) {
1991 return true;
1992 }
1993 } else if (opc == Op_SubI) {
1994 if (scaled_iv(n->in(1)) && offset_plus_k(n->in(2), true)) {
1995 return true;
1996 }
1997 if (scaled_iv(n->in(2)) && offset_plus_k(n->in(1))) {
1998 _scale *= -1;
1999 return true;
2000 }
2001 }
2002 return false;
2003 }
2004
2005 //----------------------------scaled_iv------------------------
2006 // Match: k*iv where k is a constant that's not zero
2007 bool SWPointer::scaled_iv(Node* n) {
2008 if (_scale != 0) {
2009 return false; // already found a scale
2010 }
2011 if (n == iv()) {
2012 _scale = 1;
2013 return true;
2014 }
2015 int opc = n->Opcode();
2016 if (opc == Op_MulI) {
2017 if (n->in(1) == iv() && n->in(2)->is_Con()) {
2018 _scale = n->in(2)->get_int();
2019 return true;
2020 } else if (n->in(2) == iv() && n->in(1)->is_Con()) {
2021 _scale = n->in(1)->get_int();
2022 return true;
2023 }
2024 } else if (opc == Op_LShiftI) {
2025 if (n->in(1) == iv() && n->in(2)->is_Con()) {
2026 _scale = 1 << n->in(2)->get_int();
2027 return true;
2028 }
2029 } else if (opc == Op_ConvI2L) {
2030 if (scaled_iv_plus_offset(n->in(1))) {
2031 return true;
2032 }
2033 } else if (opc == Op_LShiftL) {
2034 if (!has_iv() && _invar == NULL) {
2035 // Need to preserve the current _offset value, so
2036 // create a temporary object for this expression subtree.
2037 // Hacky, so should re-engineer the address pattern match.
2038 SWPointer tmp(this);
2039 if (tmp.scaled_iv_plus_offset(n->in(1))) {
2040 if (tmp._invar == NULL) {
2041 int mult = 1 << n->in(2)->get_int();
2042 _scale = tmp._scale * mult;
2043 _offset += tmp._offset * mult;
2044 return true;
2045 }
2046 }
2047 }
2048 }
2049 return false;
2050 }
2051
2052 //----------------------------offset_plus_k------------------------
2053 // Match: offset is (k [+/- invariant])
2054 // where k maybe zero and invariant is optional, but not both.
2055 bool SWPointer::offset_plus_k(Node* n, bool negate) {
2056 int opc = n->Opcode();
2057 if (opc == Op_ConI) {
2058 _offset += negate ? -(n->get_int()) : n->get_int();
2059 return true;
2060 } else if (opc == Op_ConL) {
2061 // Okay if value fits into an int
2062 const TypeLong* t = n->find_long_type();
2063 if (t->higher_equal(TypeLong::INT)) {
2064 jlong loff = n->get_long();
2065 jint off = (jint)loff;
2066 _offset += negate ? -off : loff;
2067 return true;
2068 }
2069 return false;
2070 }
2071 if (_invar != NULL) return false; // already have an invariant
2072 if (opc == Op_AddI) {
2073 if (n->in(2)->is_Con() && invariant(n->in(1))) {
2074 _negate_invar = negate;
2075 _invar = n->in(1);
2076 _offset += negate ? -(n->in(2)->get_int()) : n->in(2)->get_int();
2077 return true;
2078 } else if (n->in(1)->is_Con() && invariant(n->in(2))) {
2079 _offset += negate ? -(n->in(1)->get_int()) : n->in(1)->get_int();
2080 _negate_invar = negate;
2081 _invar = n->in(2);
2082 return true;
2083 }
2084 }
2085 if (opc == Op_SubI) {
2086 if (n->in(2)->is_Con() && invariant(n->in(1))) {
2087 _negate_invar = negate;
2088 _invar = n->in(1);
2089 _offset += !negate ? -(n->in(2)->get_int()) : n->in(2)->get_int();
2090 return true;
2091 } else if (n->in(1)->is_Con() && invariant(n->in(2))) {
2092 _offset += negate ? -(n->in(1)->get_int()) : n->in(1)->get_int();
2093 _negate_invar = !negate;
2094 _invar = n->in(2);
2095 return true;
2096 }
2097 }
2098 if (invariant(n)) {
2099 _negate_invar = negate;
2100 _invar = n;
2101 return true;
2102 }
2103 return false;
2104 }
2105
2106 //----------------------------print------------------------
2107 void SWPointer::print() {
2108 #ifndef PRODUCT
2109 tty->print("base: %d adr: %d scale: %d offset: %d invar: %c%d\n",
2110 _base != NULL ? _base->_idx : 0,
2111 _adr != NULL ? _adr->_idx : 0,
2112 _scale, _offset,
2113 _negate_invar?'-':'+',
2114 _invar != NULL ? _invar->_idx : 0);
2115 #endif
2116 }
2117
2118 // ========================= OrderedPair =====================
2119
2120 const OrderedPair OrderedPair::initial;
2121
2122 // ========================= SWNodeInfo =====================
2123
2124 const SWNodeInfo SWNodeInfo::initial;
2125
2126
2127 // ============================ DepGraph ===========================
2128
2129 //------------------------------make_node---------------------------
2130 // Make a new dependence graph node for an ideal node.
2131 DepMem* DepGraph::make_node(Node* node) {
2132 DepMem* m = new (_arena) DepMem(node);
2133 if (node != NULL) {
2134 assert(_map.at_grow(node->_idx) == NULL, "one init only");
2135 _map.at_put_grow(node->_idx, m);
2136 }
2137 return m;
2138 }
2139
2140 //------------------------------make_edge---------------------------
2141 // Make a new dependence graph edge from dpred -> dsucc
2142 DepEdge* DepGraph::make_edge(DepMem* dpred, DepMem* dsucc) {
2143 DepEdge* e = new (_arena) DepEdge(dpred, dsucc, dsucc->in_head(), dpred->out_head());
2144 dpred->set_out_head(e);
2145 dsucc->set_in_head(e);
2146 return e;
2147 }
2148
2149 // ========================== DepMem ========================
2150
2151 //------------------------------in_cnt---------------------------
2152 int DepMem::in_cnt() {
2153 int ct = 0;
2154 for (DepEdge* e = _in_head; e != NULL; e = e->next_in()) ct++;
2155 return ct;
2156 }
2157
2158 //------------------------------out_cnt---------------------------
2159 int DepMem::out_cnt() {
2160 int ct = 0;
2161 for (DepEdge* e = _out_head; e != NULL; e = e->next_out()) ct++;
2162 return ct;
2163 }
2164
2165 //------------------------------print-----------------------------
2166 void DepMem::print() {
2167 #ifndef PRODUCT
2168 tty->print(" DepNode %d (", _node->_idx);
2169 for (DepEdge* p = _in_head; p != NULL; p = p->next_in()) {
2170 Node* pred = p->pred()->node();
2171 tty->print(" %d", pred != NULL ? pred->_idx : 0);
2172 }
2173 tty->print(") [");
2174 for (DepEdge* s = _out_head; s != NULL; s = s->next_out()) {
2175 Node* succ = s->succ()->node();
2176 tty->print(" %d", succ != NULL ? succ->_idx : 0);
2177 }
2178 tty->print_cr(" ]");
2179 #endif
2180 }
2181
2182 // =========================== DepEdge =========================
2183
2184 //------------------------------DepPreds---------------------------
2185 void DepEdge::print() {
2186 #ifndef PRODUCT
2187 tty->print_cr("DepEdge: %d [ %d ]", _pred->node()->_idx, _succ->node()->_idx);
2188 #endif
2189 }
2190
2191 // =========================== DepPreds =========================
2192 // Iterator over predecessor edges in the dependence graph.
2193
2194 //------------------------------DepPreds---------------------------
2195 DepPreds::DepPreds(Node* n, DepGraph& dg) {
2196 _n = n;
2197 _done = false;
2198 if (_n->is_Store() || _n->is_Load()) {
2199 _next_idx = MemNode::Address;
2200 _end_idx = n->req();
2201 _dep_next = dg.dep(_n)->in_head();
2202 } else if (_n->is_Mem()) {
2203 _next_idx = 0;
2204 _end_idx = 0;
2205 _dep_next = dg.dep(_n)->in_head();
2206 } else {
2207 _next_idx = 1;
2208 _end_idx = _n->req();
2209 _dep_next = NULL;
2210 }
2211 next();
2212 }
2213
2214 //------------------------------next---------------------------
2215 void DepPreds::next() {
2216 if (_dep_next != NULL) {
2217 _current = _dep_next->pred()->node();
2218 _dep_next = _dep_next->next_in();
2219 } else if (_next_idx < _end_idx) {
2220 _current = _n->in(_next_idx++);
2221 } else {
2222 _done = true;
2223 }
2224 }
2225
2226 // =========================== DepSuccs =========================
2227 // Iterator over successor edges in the dependence graph.
2228
2229 //------------------------------DepSuccs---------------------------
2230 DepSuccs::DepSuccs(Node* n, DepGraph& dg) {
2231 _n = n;
2232 _done = false;
2233 if (_n->is_Load()) {
2234 _next_idx = 0;
2235 _end_idx = _n->outcnt();
2236 _dep_next = dg.dep(_n)->out_head();
2237 } else if (_n->is_Mem() || _n->is_Phi() && _n->bottom_type() == Type::MEMORY) {
2238 _next_idx = 0;
2239 _end_idx = 0;
2240 _dep_next = dg.dep(_n)->out_head();
2241 } else {
2242 _next_idx = 0;
2243 _end_idx = _n->outcnt();
2244 _dep_next = NULL;
2245 }
2246 next();
2247 }
2248
2249 //-------------------------------next---------------------------
2250 void DepSuccs::next() {
2251 if (_dep_next != NULL) {
2252 _current = _dep_next->succ()->node();
2253 _dep_next = _dep_next->next_out();
2254 } else if (_next_idx < _end_idx) {
2255 _current = _n->raw_out(_next_idx++);
2256 } else {
2257 _done = true;
2258 }
2259 }