1 /*
2 * Copyright (c) 2007, 2013, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 */
23
24 #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(UseSuperWord, "should be");
71 // Do vectors exist on this architecture?
72 if (Matcher::vector_width_in_bytes(T_BYTE) < 2) return;
73
74 assert(lpt->_head->is_CountedLoop(), "must be");
75 CountedLoopNode *cl = lpt->_head->as_CountedLoop();
76
77 if (!cl->is_valid_counted_loop()) return; // skip malformed counted loop
78
79 if (!cl->is_main_loop() ) return; // skip normal, pre, and post loops
80
81 // Check for no control flow in body (other than exit)
82 Node *cl_exit = cl->loopexit();
83 if (cl_exit->in(0) != lpt->_head) return;
84
85 // Make sure the are no extra control users of the loop backedge
86 if (cl->back_control()->outcnt() != 1) {
87 return;
88 }
89
90 // Check for pre-loop ending with CountedLoopEnd(Bool(Cmp(x,Opaque1(limit))))
91 CountedLoopEndNode* pre_end = get_pre_loop_end(cl);
92 if (pre_end == NULL) return;
93 Node *pre_opaq1 = pre_end->limit();
94 if (pre_opaq1->Opcode() != Op_Opaque1) return;
95
96 init(); // initialize data structures
97
98 set_lpt(lpt);
99 set_lp(cl);
100
101 // For now, define one block which is the entire loop body
102 set_bb(cl);
103
104 assert(_packset.length() == 0, "packset must be empty");
105 SLP_extract();
106 }
107
108 //------------------------------SLP_extract---------------------------
109 // Extract the superword level parallelism
110 //
111 // 1) A reverse post-order of nodes in the block is constructed. By scanning
112 // this list from first to last, all definitions are visited before their uses.
113 //
114 // 2) A point-to-point dependence graph is constructed between memory references.
115 // This simplies the upcoming "independence" checker.
116 //
117 // 3) The maximum depth in the node graph from the beginning of the block
118 // to each node is computed. This is used to prune the graph search
119 // in the independence checker.
120 //
121 // 4) For integer types, the necessary bit width is propagated backwards
122 // from stores to allow packed operations on byte, char, and short
123 // integers. This reverses the promotion to type "int" that javac
124 // did for operations like: char c1,c2,c3; c1 = c2 + c3.
125 //
126 // 5) One of the memory references is picked to be an aligned vector reference.
127 // The pre-loop trip count is adjusted to align this reference in the
128 // unrolled body.
129 //
130 // 6) The initial set of pack pairs is seeded with memory references.
131 //
132 // 7) The set of pack pairs is extended by following use->def and def->use links.
133 //
134 // 8) The pairs are combined into vector sized packs.
135 //
136 // 9) Reorder the memory slices to co-locate members of the memory packs.
137 //
138 // 10) Generate ideal vector nodes for the final set of packs and where necessary,
139 // inserting scalar promotion, vector creation from multiple scalars, and
140 // extraction of scalar values from vectors.
141 //
142 void SuperWord::SLP_extract() {
143
144 // Ready the block
145
146 if (!construct_bb())
147 return; // Exit if no interesting nodes or complex graph.
148
149 dependence_graph();
150
151 compute_max_depth();
152
153 compute_vector_element_type();
154
155 // Attempt vectorization
156
157 find_adjacent_refs();
158
159 extend_packlist();
160
161 combine_packs();
162
163 construct_my_pack_map();
164
165 filter_packs();
166
167 schedule();
168
169 output();
170 }
171
172 //------------------------------find_adjacent_refs---------------------------
173 // Find the adjacent memory references and create pack pairs for them.
174 // This is the initial set of packs that will then be extended by
175 // following use->def and def->use links. The align positions are
176 // assigned relative to the reference "align_to_ref"
177 void SuperWord::find_adjacent_refs() {
178 // Get list of memory operations
179 Node_List memops;
180 for (int i = 0; i < _block.length(); i++) {
181 Node* n = _block.at(i);
182 if (n->is_Mem() && !n->is_LoadStore() && in_bb(n) &&
183 is_java_primitive(n->as_Mem()->memory_type())) {
184 int align = memory_alignment(n->as_Mem(), 0);
185 if (align != bottom_align) {
186 memops.push(n);
187 }
188 }
189 }
190
191 Node_List align_to_refs;
192 int best_iv_adjustment = 0;
193 MemNode* best_align_to_mem_ref = NULL;
194
195 while (memops.size() != 0) {
196 // Find a memory reference to align to.
197 MemNode* mem_ref = find_align_to_ref(memops);
198 if (mem_ref == NULL) break;
199 align_to_refs.push(mem_ref);
200 int iv_adjustment = get_iv_adjustment(mem_ref);
201
202 if (best_align_to_mem_ref == NULL) {
203 // Set memory reference which is the best from all memory operations
204 // to be used for alignment. The pre-loop trip count is modified to align
205 // this reference to a vector-aligned address.
206 best_align_to_mem_ref = mem_ref;
207 best_iv_adjustment = iv_adjustment;
208 }
209
210 SWPointer align_to_ref_p(mem_ref, this);
211 // Set alignment relative to "align_to_ref" for all related memory operations.
212 for (int i = memops.size() - 1; i >= 0; i--) {
213 MemNode* s = memops.at(i)->as_Mem();
214 if (isomorphic(s, mem_ref)) {
215 SWPointer p2(s, this);
216 if (p2.comparable(align_to_ref_p)) {
217 int align = memory_alignment(s, iv_adjustment);
218 set_alignment(s, align);
219 }
220 }
221 }
222
223 // Create initial pack pairs of memory operations for which
224 // alignment is set and vectors will be aligned.
225 bool create_pack = true;
226 if (memory_alignment(mem_ref, best_iv_adjustment) == 0) {
227 if (!Matcher::misaligned_vectors_ok()) {
228 int vw = vector_width(mem_ref);
229 int vw_best = vector_width(best_align_to_mem_ref);
230 if (vw > vw_best) {
231 // Do not vectorize a memory access with more elements per vector
232 // if unaligned memory access is not allowed because number of
233 // iterations in pre-loop will be not enough to align it.
234 create_pack = false;
235 }
236 }
237 } else {
238 if (same_velt_type(mem_ref, best_align_to_mem_ref)) {
239 // Can't allow vectorization of unaligned memory accesses with the
240 // same type since it could be overlapped accesses to the same array.
241 create_pack = false;
242 } else {
243 // Allow independent (different type) unaligned memory operations
244 // if HW supports them.
245 if (!Matcher::misaligned_vectors_ok()) {
246 create_pack = false;
247 } else {
248 // Check if packs of the same memory type but
249 // with a different alignment were created before.
250 for (uint i = 0; i < align_to_refs.size(); i++) {
251 MemNode* mr = align_to_refs.at(i)->as_Mem();
252 if (same_velt_type(mr, mem_ref) &&
253 memory_alignment(mr, iv_adjustment) != 0)
254 create_pack = false;
255 }
256 }
257 }
258 }
259 if (create_pack) {
260 for (uint i = 0; i < memops.size(); i++) {
261 Node* s1 = memops.at(i);
262 int align = alignment(s1);
263 if (align == top_align) continue;
264 for (uint j = 0; j < memops.size(); j++) {
265 Node* s2 = memops.at(j);
266 if (alignment(s2) == top_align) continue;
267 if (s1 != s2 && are_adjacent_refs(s1, s2)) {
268 if (stmts_can_pack(s1, s2, align)) {
269 Node_List* pair = new Node_List();
270 pair->push(s1);
271 pair->push(s2);
272 _packset.append(pair);
273 }
274 }
275 }
276 }
277 } else { // Don't create unaligned pack
278 // First, remove remaining memory ops of the same type from the list.
279 for (int i = memops.size() - 1; i >= 0; i--) {
280 MemNode* s = memops.at(i)->as_Mem();
281 if (same_velt_type(s, mem_ref)) {
282 memops.remove(i);
283 }
284 }
285
286 // Second, remove already constructed packs of the same type.
287 for (int i = _packset.length() - 1; i >= 0; i--) {
288 Node_List* p = _packset.at(i);
289 MemNode* s = p->at(0)->as_Mem();
290 if (same_velt_type(s, mem_ref)) {
291 remove_pack_at(i);
292 }
293 }
294
295 // If needed find the best memory reference for loop alignment again.
296 if (same_velt_type(mem_ref, best_align_to_mem_ref)) {
297 // Put memory ops from remaining packs back on memops list for
298 // the best alignment search.
299 uint orig_msize = memops.size();
300 for (int i = 0; i < _packset.length(); i++) {
301 Node_List* p = _packset.at(i);
302 MemNode* s = p->at(0)->as_Mem();
303 assert(!same_velt_type(s, mem_ref), "sanity");
304 memops.push(s);
305 }
306 MemNode* best_align_to_mem_ref = find_align_to_ref(memops);
307 if (best_align_to_mem_ref == NULL) break;
308 best_iv_adjustment = get_iv_adjustment(best_align_to_mem_ref);
309 // Restore list.
310 while (memops.size() > orig_msize)
311 (void)memops.pop();
312 }
313 } // unaligned memory accesses
314
315 // Remove used mem nodes.
316 for (int i = memops.size() - 1; i >= 0; i--) {
317 MemNode* m = memops.at(i)->as_Mem();
318 if (alignment(m) != top_align) {
319 memops.remove(i);
320 }
321 }
322
323 } // while (memops.size() != 0
324 set_align_to_ref(best_align_to_mem_ref);
325
326 #ifndef PRODUCT
327 if (TraceSuperWord) {
328 tty->print_cr("\nAfter find_adjacent_refs");
329 print_packset();
330 }
331 #endif
332 }
333
334 //------------------------------find_align_to_ref---------------------------
335 // Find a memory reference to align the loop induction variable to.
336 // Looks first at stores then at loads, looking for a memory reference
337 // with the largest number of references similar to it.
338 MemNode* SuperWord::find_align_to_ref(Node_List &memops) {
339 GrowableArray<int> cmp_ct(arena(), memops.size(), memops.size(), 0);
340
341 // Count number of comparable memory ops
342 for (uint i = 0; i < memops.size(); i++) {
343 MemNode* s1 = memops.at(i)->as_Mem();
344 SWPointer p1(s1, this);
345 // Discard if pre loop can't align this reference
346 if (!ref_is_alignable(p1)) {
347 *cmp_ct.adr_at(i) = 0;
348 continue;
349 }
350 for (uint j = i+1; j < memops.size(); j++) {
351 MemNode* s2 = memops.at(j)->as_Mem();
352 if (isomorphic(s1, s2)) {
353 SWPointer p2(s2, this);
354 if (p1.comparable(p2)) {
355 (*cmp_ct.adr_at(i))++;
356 (*cmp_ct.adr_at(j))++;
357 }
358 }
359 }
360 }
361
362 // Find Store (or Load) with the greatest number of "comparable" references,
363 // biggest vector size, smallest data size and smallest iv offset.
364 int max_ct = 0;
365 int max_vw = 0;
366 int max_idx = -1;
367 int min_size = max_jint;
368 int min_iv_offset = max_jint;
369 for (uint j = 0; j < memops.size(); j++) {
370 MemNode* s = memops.at(j)->as_Mem();
371 if (s->is_Store()) {
372 int vw = vector_width_in_bytes(s);
373 assert(vw > 1, "sanity");
374 SWPointer p(s, this);
375 if (cmp_ct.at(j) > max_ct ||
376 cmp_ct.at(j) == max_ct &&
377 (vw > max_vw ||
378 vw == max_vw &&
379 (data_size(s) < min_size ||
380 data_size(s) == min_size &&
381 (p.offset_in_bytes() < min_iv_offset)))) {
382 max_ct = cmp_ct.at(j);
383 max_vw = vw;
384 max_idx = j;
385 min_size = data_size(s);
386 min_iv_offset = p.offset_in_bytes();
387 }
388 }
389 }
390 // If no stores, look at loads
391 if (max_ct == 0) {
392 for (uint j = 0; j < memops.size(); j++) {
393 MemNode* s = memops.at(j)->as_Mem();
394 if (s->is_Load()) {
395 int vw = vector_width_in_bytes(s);
396 assert(vw > 1, "sanity");
397 SWPointer p(s, this);
398 if (cmp_ct.at(j) > max_ct ||
399 cmp_ct.at(j) == max_ct &&
400 (vw > max_vw ||
401 vw == max_vw &&
402 (data_size(s) < min_size ||
403 data_size(s) == min_size &&
404 (p.offset_in_bytes() < min_iv_offset)))) {
405 max_ct = cmp_ct.at(j);
406 max_vw = vw;
407 max_idx = j;
408 min_size = data_size(s);
409 min_iv_offset = p.offset_in_bytes();
410 }
411 }
412 }
413 }
414
415 #ifdef ASSERT
416 if (TraceSuperWord && Verbose) {
417 tty->print_cr("\nVector memops after find_align_to_refs");
418 for (uint i = 0; i < memops.size(); i++) {
419 MemNode* s = memops.at(i)->as_Mem();
420 s->dump();
421 }
422 }
423 #endif
424
425 if (max_ct > 0) {
426 #ifdef ASSERT
427 if (TraceSuperWord) {
428 tty->print("\nVector align to node: ");
429 memops.at(max_idx)->as_Mem()->dump();
430 }
431 #endif
432 return memops.at(max_idx)->as_Mem();
433 }
434 return NULL;
435 }
436
437 //------------------------------ref_is_alignable---------------------------
438 // Can the preloop align the reference to position zero in the vector?
439 bool SuperWord::ref_is_alignable(SWPointer& p) {
440 if (!p.has_iv()) {
441 return true; // no induction variable
442 }
443 CountedLoopEndNode* pre_end = get_pre_loop_end(lp()->as_CountedLoop());
444 assert(pre_end != NULL, "we must have a correct pre-loop");
445 assert(pre_end->stride_is_con(), "pre loop stride is constant");
446 int preloop_stride = pre_end->stride_con();
447
448 int span = preloop_stride * p.scale_in_bytes();
449
450 // Stride one accesses are alignable.
451 if (ABS(span) == p.memory_size())
452 return true;
453
454 // If initial offset from start of object is computable,
455 // compute alignment within the vector.
456 int vw = vector_width_in_bytes(p.mem());
457 assert(vw > 1, "sanity");
458 if (vw % span == 0) {
459 Node* init_nd = pre_end->init_trip();
460 if (init_nd->is_Con() && p.invar() == NULL) {
461 int init = init_nd->bottom_type()->is_int()->get_con();
462
463 int init_offset = init * p.scale_in_bytes() + p.offset_in_bytes();
464 assert(init_offset >= 0, "positive offset from object start");
465
466 if (span > 0) {
467 return (vw - (init_offset % vw)) % span == 0;
468 } else {
469 assert(span < 0, "nonzero stride * scale");
470 return (init_offset % vw) % -span == 0;
471 }
472 }
473 }
474 return false;
475 }
476
477 //---------------------------get_iv_adjustment---------------------------
478 // Calculate loop's iv adjustment for this memory ops.
479 int SuperWord::get_iv_adjustment(MemNode* mem_ref) {
480 SWPointer align_to_ref_p(mem_ref, this);
481 int offset = align_to_ref_p.offset_in_bytes();
482 int scale = align_to_ref_p.scale_in_bytes();
483 int vw = vector_width_in_bytes(mem_ref);
484 assert(vw > 1, "sanity");
485 int stride_sign = (scale * iv_stride()) > 0 ? 1 : -1;
486 // At least one iteration is executed in pre-loop by default. As result
487 // several iterations are needed to align memory operations in main-loop even
488 // if offset is 0.
489 int iv_adjustment_in_bytes = (stride_sign * vw - (offset % vw));
490 int elt_size = align_to_ref_p.memory_size();
491 assert(((ABS(iv_adjustment_in_bytes) % elt_size) == 0),
492 err_msg_res("(%d) should be divisible by (%d)", iv_adjustment_in_bytes, elt_size));
493 int iv_adjustment = iv_adjustment_in_bytes/elt_size;
494
495 #ifndef PRODUCT
496 if (TraceSuperWord)
497 tty->print_cr("\noffset = %d iv_adjust = %d elt_size = %d scale = %d iv_stride = %d vect_size %d",
498 offset, iv_adjustment, elt_size, scale, iv_stride(), vw);
499 #endif
500 return iv_adjustment;
501 }
502
503 //---------------------------dependence_graph---------------------------
504 // Construct dependency graph.
505 // Add dependence edges to load/store nodes for memory dependence
506 // A.out()->DependNode.in(1) and DependNode.out()->B.prec(x)
507 void SuperWord::dependence_graph() {
508 // First, assign a dependence node to each memory node
509 for (int i = 0; i < _block.length(); i++ ) {
510 Node *n = _block.at(i);
511 if (n->is_Mem() || n->is_Phi() && n->bottom_type() == Type::MEMORY) {
512 _dg.make_node(n);
513 }
514 }
515
516 // For each memory slice, create the dependences
517 for (int i = 0; i < _mem_slice_head.length(); i++) {
518 Node* n = _mem_slice_head.at(i);
519 Node* n_tail = _mem_slice_tail.at(i);
520
521 // Get slice in predecessor order (last is first)
522 mem_slice_preds(n_tail, n, _nlist);
523
524 // Make the slice dependent on the root
525 DepMem* slice = _dg.dep(n);
526 _dg.make_edge(_dg.root(), slice);
527
528 // Create a sink for the slice
529 DepMem* slice_sink = _dg.make_node(NULL);
530 _dg.make_edge(slice_sink, _dg.tail());
531
532 // Now visit each pair of memory ops, creating the edges
533 for (int j = _nlist.length() - 1; j >= 0 ; j--) {
534 Node* s1 = _nlist.at(j);
535
536 // If no dependency yet, use slice
537 if (_dg.dep(s1)->in_cnt() == 0) {
538 _dg.make_edge(slice, s1);
539 }
540 SWPointer p1(s1->as_Mem(), this);
541 bool sink_dependent = true;
542 for (int k = j - 1; k >= 0; k--) {
543 Node* s2 = _nlist.at(k);
544 if (s1->is_Load() && s2->is_Load())
545 continue;
546 SWPointer p2(s2->as_Mem(), this);
547
548 int cmp = p1.cmp(p2);
549 if (SuperWordRTDepCheck &&
550 p1.base() != p2.base() && p1.valid() && p2.valid()) {
551 // Create a runtime check to disambiguate
552 OrderedPair pp(p1.base(), p2.base());
553 _disjoint_ptrs.append_if_missing(pp);
554 } else if (!SWPointer::not_equal(cmp)) {
555 // Possibly same address
556 _dg.make_edge(s1, s2);
557 sink_dependent = false;
558 }
559 }
560 if (sink_dependent) {
561 _dg.make_edge(s1, slice_sink);
562 }
563 }
564 #ifndef PRODUCT
565 if (TraceSuperWord) {
566 tty->print_cr("\nDependence graph for slice: %d", n->_idx);
567 for (int q = 0; q < _nlist.length(); q++) {
568 _dg.print(_nlist.at(q));
569 }
570 tty->cr();
571 }
572 #endif
573 _nlist.clear();
574 }
575
576 #ifndef PRODUCT
577 if (TraceSuperWord) {
578 tty->print_cr("\ndisjoint_ptrs: %s", _disjoint_ptrs.length() > 0 ? "" : "NONE");
579 for (int r = 0; r < _disjoint_ptrs.length(); r++) {
580 _disjoint_ptrs.at(r).print();
581 tty->cr();
582 }
583 tty->cr();
584 }
585 #endif
586 }
587
588 //---------------------------mem_slice_preds---------------------------
589 // Return a memory slice (node list) in predecessor order starting at "start"
590 void SuperWord::mem_slice_preds(Node* start, Node* stop, GrowableArray<Node*> &preds) {
591 assert(preds.length() == 0, "start empty");
592 Node* n = start;
593 Node* prev = NULL;
594 while (true) {
595 assert(in_bb(n), "must be in block");
596 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
597 Node* out = n->fast_out(i);
598 if (out->is_Load()) {
599 if (in_bb(out)) {
600 preds.push(out);
601 }
602 } else {
603 // FIXME
604 if (out->is_MergeMem() && !in_bb(out)) {
605 // Either unrolling is causing a memory edge not to disappear,
606 // or need to run igvn.optimize() again before SLP
607 } else if (out->is_Phi() && out->bottom_type() == Type::MEMORY && !in_bb(out)) {
608 // Ditto. Not sure what else to check further.
609 } else if (out->Opcode() == Op_StoreCM && out->in(MemNode::OopStore) == n) {
610 // StoreCM has an input edge used as a precedence edge.
611 // Maybe an issue when oop stores are vectorized.
612 } else {
613 assert(out == prev || prev == NULL, "no branches off of store slice");
614 }
615 }
616 }
617 if (n == stop) break;
618 preds.push(n);
619 prev = n;
620 assert(n->is_Mem(), err_msg_res("unexpected node %s", n->Name()));
621 n = n->in(MemNode::Memory);
622 }
623 }
624
625 //------------------------------stmts_can_pack---------------------------
626 // Can s1 and s2 be in a pack with s1 immediately preceding s2 and
627 // s1 aligned at "align"
628 bool SuperWord::stmts_can_pack(Node* s1, Node* s2, int align) {
629
630 // Do not use superword for non-primitives
631 BasicType bt1 = velt_basic_type(s1);
632 BasicType bt2 = velt_basic_type(s2);
633 if(!is_java_primitive(bt1) || !is_java_primitive(bt2))
634 return false;
635 if (Matcher::max_vector_size(bt1) < 2) {
636 return false; // No vectors for this type
637 }
638
639 if (isomorphic(s1, s2)) {
640 if (independent(s1, s2)) {
641 if (!exists_at(s1, 0) && !exists_at(s2, 1)) {
642 if (!s1->is_Mem() || are_adjacent_refs(s1, s2)) {
643 int s1_align = alignment(s1);
644 int s2_align = alignment(s2);
645 if (s1_align == top_align || s1_align == align) {
646 if (s2_align == top_align || s2_align == align + data_size(s1)) {
647 return true;
648 }
649 }
650 }
651 }
652 }
653 }
654 return false;
655 }
656
657 //------------------------------exists_at---------------------------
658 // Does s exist in a pack at position pos?
659 bool SuperWord::exists_at(Node* s, uint pos) {
660 for (int i = 0; i < _packset.length(); i++) {
661 Node_List* p = _packset.at(i);
662 if (p->at(pos) == s) {
663 return true;
664 }
665 }
666 return false;
667 }
668
669 //------------------------------are_adjacent_refs---------------------------
670 // Is s1 immediately before s2 in memory?
671 bool SuperWord::are_adjacent_refs(Node* s1, Node* s2) {
672 if (!s1->is_Mem() || !s2->is_Mem()) return false;
673 if (!in_bb(s1) || !in_bb(s2)) return false;
674
675 // Do not use superword for non-primitives
676 if (!is_java_primitive(s1->as_Mem()->memory_type()) ||
677 !is_java_primitive(s2->as_Mem()->memory_type())) {
678 return false;
679 }
680
681 // FIXME - co_locate_pack fails on Stores in different mem-slices, so
682 // only pack memops that are in the same alias set until that's fixed.
683 if (_phase->C->get_alias_index(s1->as_Mem()->adr_type()) !=
684 _phase->C->get_alias_index(s2->as_Mem()->adr_type()))
685 return false;
686 SWPointer p1(s1->as_Mem(), this);
687 SWPointer p2(s2->as_Mem(), this);
688 if (p1.base() != p2.base() || !p1.comparable(p2)) return false;
689 int diff = p2.offset_in_bytes() - p1.offset_in_bytes();
690 return diff == data_size(s1);
691 }
692
693 //------------------------------isomorphic---------------------------
694 // Are s1 and s2 similar?
695 bool SuperWord::isomorphic(Node* s1, Node* s2) {
696 if (s1->Opcode() != s2->Opcode()) return false;
697 if (s1->req() != s2->req()) return false;
698 if (s1->in(0) != s2->in(0)) return false;
699 if (!same_velt_type(s1, s2)) return false;
700 return true;
701 }
702
703 //------------------------------independent---------------------------
704 // Is there no data path from s1 to s2 or s2 to s1?
705 bool SuperWord::independent(Node* s1, Node* s2) {
706 // assert(s1->Opcode() == s2->Opcode(), "check isomorphic first");
707 int d1 = depth(s1);
708 int d2 = depth(s2);
709 if (d1 == d2) return s1 != s2;
710 Node* deep = d1 > d2 ? s1 : s2;
711 Node* shallow = d1 > d2 ? s2 : s1;
712
713 visited_clear();
714
715 return independent_path(shallow, deep);
716 }
717
718 //------------------------------independent_path------------------------------
719 // Helper for independent
720 bool SuperWord::independent_path(Node* shallow, Node* deep, uint dp) {
721 if (dp >= 1000) return false; // stop deep recursion
722 visited_set(deep);
723 int shal_depth = depth(shallow);
724 assert(shal_depth <= depth(deep), "must be");
725 for (DepPreds preds(deep, _dg); !preds.done(); preds.next()) {
726 Node* pred = preds.current();
727 if (in_bb(pred) && !visited_test(pred)) {
728 if (shallow == pred) {
729 return false;
730 }
731 if (shal_depth < depth(pred) && !independent_path(shallow, pred, dp+1)) {
732 return false;
733 }
734 }
735 }
736 return true;
737 }
738
739 //------------------------------set_alignment---------------------------
740 void SuperWord::set_alignment(Node* s1, Node* s2, int align) {
741 set_alignment(s1, align);
742 if (align == top_align || align == bottom_align) {
743 set_alignment(s2, align);
744 } else {
745 set_alignment(s2, align + data_size(s1));
746 }
747 }
748
749 //------------------------------data_size---------------------------
750 int SuperWord::data_size(Node* s) {
751 int bsize = type2aelembytes(velt_basic_type(s));
752 assert(bsize != 0, "valid size");
753 return bsize;
754 }
755
756 //------------------------------extend_packlist---------------------------
757 // Extend packset by following use->def and def->use links from pack members.
758 void SuperWord::extend_packlist() {
759 bool changed;
760 do {
761 changed = false;
762 for (int i = 0; i < _packset.length(); i++) {
763 Node_List* p = _packset.at(i);
764 changed |= follow_use_defs(p);
765 changed |= follow_def_uses(p);
766 }
767 } while (changed);
768
769 #ifndef PRODUCT
770 if (TraceSuperWord) {
771 tty->print_cr("\nAfter extend_packlist");
772 print_packset();
773 }
774 #endif
775 }
776
777 //------------------------------follow_use_defs---------------------------
778 // Extend the packset by visiting operand definitions of nodes in pack p
779 bool SuperWord::follow_use_defs(Node_List* p) {
780 assert(p->size() == 2, "just checking");
781 Node* s1 = p->at(0);
782 Node* s2 = p->at(1);
783 assert(s1->req() == s2->req(), "just checking");
784 assert(alignment(s1) + data_size(s1) == alignment(s2), "just checking");
785
786 if (s1->is_Load()) return false;
787
788 int align = alignment(s1);
789 bool changed = false;
790 int start = s1->is_Store() ? MemNode::ValueIn : 1;
791 int end = s1->is_Store() ? MemNode::ValueIn+1 : s1->req();
792 for (int j = start; j < end; j++) {
793 Node* t1 = s1->in(j);
794 Node* t2 = s2->in(j);
795 if (!in_bb(t1) || !in_bb(t2))
796 continue;
797 if (stmts_can_pack(t1, t2, align)) {
798 if (est_savings(t1, t2) >= 0) {
799 Node_List* pair = new Node_List();
800 pair->push(t1);
801 pair->push(t2);
802 _packset.append(pair);
803 set_alignment(t1, t2, align);
804 changed = true;
805 }
806 }
807 }
808 return changed;
809 }
810
811 //------------------------------follow_def_uses---------------------------
812 // Extend the packset by visiting uses of nodes in pack p
813 bool SuperWord::follow_def_uses(Node_List* p) {
814 bool changed = false;
815 Node* s1 = p->at(0);
816 Node* s2 = p->at(1);
817 assert(p->size() == 2, "just checking");
818 assert(s1->req() == s2->req(), "just checking");
819 assert(alignment(s1) + data_size(s1) == alignment(s2), "just checking");
820
821 if (s1->is_Store()) return false;
822
823 int align = alignment(s1);
824 int savings = -1;
825 Node* u1 = NULL;
826 Node* u2 = NULL;
827 for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
828 Node* t1 = s1->fast_out(i);
829 if (!in_bb(t1)) continue;
830 for (DUIterator_Fast jmax, j = s2->fast_outs(jmax); j < jmax; j++) {
831 Node* t2 = s2->fast_out(j);
832 if (!in_bb(t2)) continue;
833 if (!opnd_positions_match(s1, t1, s2, t2))
834 continue;
835 if (stmts_can_pack(t1, t2, align)) {
836 int my_savings = est_savings(t1, t2);
837 if (my_savings > savings) {
838 savings = my_savings;
839 u1 = t1;
840 u2 = t2;
841 }
842 }
843 }
844 }
845 if (savings >= 0) {
846 Node_List* pair = new Node_List();
847 pair->push(u1);
848 pair->push(u2);
849 _packset.append(pair);
850 set_alignment(u1, u2, align);
851 changed = true;
852 }
853 return changed;
854 }
855
856 //---------------------------opnd_positions_match-------------------------
857 // Is the use of d1 in u1 at the same operand position as d2 in u2?
858 bool SuperWord::opnd_positions_match(Node* d1, Node* u1, Node* d2, Node* u2) {
859 uint ct = u1->req();
860 if (ct != u2->req()) return false;
861 uint i1 = 0;
862 uint i2 = 0;
863 do {
864 for (i1++; i1 < ct; i1++) if (u1->in(i1) == d1) break;
865 for (i2++; i2 < ct; i2++) if (u2->in(i2) == d2) break;
866 if (i1 != i2) {
867 if ((i1 == (3-i2)) && (u2->is_Add() || u2->is_Mul())) {
868 // Further analysis relies on operands position matching.
869 u2->swap_edges(i1, i2);
870 } else {
871 return false;
872 }
873 }
874 } while (i1 < ct);
875 return true;
876 }
877
878 //------------------------------est_savings---------------------------
879 // Estimate the savings from executing s1 and s2 as a pack
880 int SuperWord::est_savings(Node* s1, Node* s2) {
881 int save_in = 2 - 1; // 2 operations per instruction in packed form
882
883 // inputs
884 for (uint i = 1; i < s1->req(); i++) {
885 Node* x1 = s1->in(i);
886 Node* x2 = s2->in(i);
887 if (x1 != x2) {
888 if (are_adjacent_refs(x1, x2)) {
889 save_in += adjacent_profit(x1, x2);
890 } else if (!in_packset(x1, x2)) {
891 save_in -= pack_cost(2);
892 } else {
893 save_in += unpack_cost(2);
894 }
895 }
896 }
897
898 // uses of result
899 uint ct = 0;
900 int save_use = 0;
901 for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
902 Node* s1_use = s1->fast_out(i);
903 for (int j = 0; j < _packset.length(); j++) {
904 Node_List* p = _packset.at(j);
905 if (p->at(0) == s1_use) {
906 for (DUIterator_Fast kmax, k = s2->fast_outs(kmax); k < kmax; k++) {
907 Node* s2_use = s2->fast_out(k);
908 if (p->at(p->size()-1) == s2_use) {
909 ct++;
910 if (are_adjacent_refs(s1_use, s2_use)) {
911 save_use += adjacent_profit(s1_use, s2_use);
912 }
913 }
914 }
915 }
916 }
917 }
918
919 if (ct < s1->outcnt()) save_use += unpack_cost(1);
920 if (ct < s2->outcnt()) save_use += unpack_cost(1);
921
922 return MAX2(save_in, save_use);
923 }
924
925 //------------------------------costs---------------------------
926 int SuperWord::adjacent_profit(Node* s1, Node* s2) { return 2; }
927 int SuperWord::pack_cost(int ct) { return ct; }
928 int SuperWord::unpack_cost(int ct) { return ct; }
929
930 //------------------------------combine_packs---------------------------
931 // Combine packs A and B with A.last == B.first into A.first..,A.last,B.second,..B.last
932 void SuperWord::combine_packs() {
933 bool changed = true;
934 // Combine packs regardless max vector size.
935 while (changed) {
936 changed = false;
937 for (int i = 0; i < _packset.length(); i++) {
938 Node_List* p1 = _packset.at(i);
939 if (p1 == NULL) continue;
940 for (int j = 0; j < _packset.length(); j++) {
941 Node_List* p2 = _packset.at(j);
942 if (p2 == NULL) continue;
943 if (i == j) continue;
944 if (p1->at(p1->size()-1) == p2->at(0)) {
945 for (uint k = 1; k < p2->size(); k++) {
946 p1->push(p2->at(k));
947 }
948 _packset.at_put(j, NULL);
949 changed = true;
950 }
951 }
952 }
953 }
954
955 // Split packs which have size greater then max vector size.
956 for (int i = 0; i < _packset.length(); i++) {
957 Node_List* p1 = _packset.at(i);
958 if (p1 != NULL) {
959 BasicType bt = velt_basic_type(p1->at(0));
960 uint max_vlen = Matcher::max_vector_size(bt); // Max elements in vector
961 assert(is_power_of_2(max_vlen), "sanity");
962 uint psize = p1->size();
963 if (!is_power_of_2(psize)) {
964 // Skip pack which can't be vector.
965 // case1: for(...) { a[i] = i; } elements values are different (i+x)
966 // case2: for(...) { a[i] = b[i+1]; } can't align both, load and store
967 _packset.at_put(i, NULL);
968 continue;
969 }
970 if (psize > max_vlen) {
971 Node_List* pack = new Node_List();
972 for (uint j = 0; j < psize; j++) {
973 pack->push(p1->at(j));
974 if (pack->size() >= max_vlen) {
975 assert(is_power_of_2(pack->size()), "sanity");
976 _packset.append(pack);
977 pack = new Node_List();
978 }
979 }
980 _packset.at_put(i, NULL);
981 }
982 }
983 }
984
985 // Compress list.
986 for (int i = _packset.length() - 1; i >= 0; i--) {
987 Node_List* p1 = _packset.at(i);
988 if (p1 == NULL) {
989 _packset.remove_at(i);
990 }
991 }
992
993 #ifndef PRODUCT
994 if (TraceSuperWord) {
995 tty->print_cr("\nAfter combine_packs");
996 print_packset();
997 }
998 #endif
999 }
1000
1001 //-----------------------------construct_my_pack_map--------------------------
1002 // Construct the map from nodes to packs. Only valid after the
1003 // point where a node is only in one pack (after combine_packs).
1004 void SuperWord::construct_my_pack_map() {
1005 Node_List* rslt = NULL;
1006 for (int i = 0; i < _packset.length(); i++) {
1007 Node_List* p = _packset.at(i);
1008 for (uint j = 0; j < p->size(); j++) {
1009 Node* s = p->at(j);
1010 assert(my_pack(s) == NULL, "only in one pack");
1011 set_my_pack(s, p);
1012 }
1013 }
1014 }
1015
1016 //------------------------------filter_packs---------------------------
1017 // Remove packs that are not implemented or not profitable.
1018 void SuperWord::filter_packs() {
1019
1020 // Remove packs that are not implemented
1021 for (int i = _packset.length() - 1; i >= 0; i--) {
1022 Node_List* pk = _packset.at(i);
1023 bool impl = implemented(pk);
1024 if (!impl) {
1025 #ifndef PRODUCT
1026 if (TraceSuperWord && Verbose) {
1027 tty->print_cr("Unimplemented");
1028 pk->at(0)->dump();
1029 }
1030 #endif
1031 remove_pack_at(i);
1032 }
1033 }
1034
1035 // Remove packs that are not profitable
1036 bool changed;
1037 do {
1038 changed = false;
1039 for (int i = _packset.length() - 1; i >= 0; i--) {
1040 Node_List* pk = _packset.at(i);
1041 bool prof = profitable(pk);
1042 if (!prof) {
1043 #ifndef PRODUCT
1044 if (TraceSuperWord && Verbose) {
1045 tty->print_cr("Unprofitable");
1046 pk->at(0)->dump();
1047 }
1048 #endif
1049 remove_pack_at(i);
1050 changed = true;
1051 }
1052 }
1053 } while (changed);
1054
1055 #ifndef PRODUCT
1056 if (TraceSuperWord) {
1057 tty->print_cr("\nAfter filter_packs");
1058 print_packset();
1059 tty->cr();
1060 }
1061 #endif
1062 }
1063
1064 //------------------------------implemented---------------------------
1065 // Can code be generated for pack p?
1066 bool SuperWord::implemented(Node_List* p) {
1067 Node* p0 = p->at(0);
1068 return VectorNode::implemented(p0->Opcode(), p->size(), velt_basic_type(p0));
1069 }
1070
1071 //------------------------------same_inputs--------------------------
1072 // For pack p, are all idx operands the same?
1073 static bool same_inputs(Node_List* p, int idx) {
1074 Node* p0 = p->at(0);
1075 uint vlen = p->size();
1076 Node* p0_def = p0->in(idx);
1077 for (uint i = 1; i < vlen; i++) {
1078 Node* pi = p->at(i);
1079 Node* pi_def = pi->in(idx);
1080 if (p0_def != pi_def)
1081 return false;
1082 }
1083 return true;
1084 }
1085
1086 //------------------------------profitable---------------------------
1087 // For pack p, are all operands and all uses (with in the block) vector?
1088 bool SuperWord::profitable(Node_List* p) {
1089 Node* p0 = p->at(0);
1090 uint start, end;
1091 VectorNode::vector_operands(p0, &start, &end);
1092
1093 // Return false if some inputs are not vectors or vectors with different
1094 // size or alignment.
1095 // Also, for now, return false if not scalar promotion case when inputs are
1096 // the same. Later, implement PackNode and allow differing, non-vector inputs
1097 // (maybe just the ones from outside the block.)
1098 for (uint i = start; i < end; i++) {
1099 if (!is_vector_use(p0, i))
1100 return false;
1101 }
1102 if (VectorNode::is_shift(p0)) {
1103 // For now, return false if shift count is vector or not scalar promotion
1104 // case (different shift counts) because it is not supported yet.
1105 Node* cnt = p0->in(2);
1106 Node_List* cnt_pk = my_pack(cnt);
1107 if (cnt_pk != NULL)
1108 return false;
1109 if (!same_inputs(p, 2))
1110 return false;
1111 }
1112 if (!p0->is_Store()) {
1113 // For now, return false if not all uses are vector.
1114 // Later, implement ExtractNode and allow non-vector uses (maybe
1115 // just the ones outside the block.)
1116 for (uint i = 0; i < p->size(); i++) {
1117 Node* def = p->at(i);
1118 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
1119 Node* use = def->fast_out(j);
1120 for (uint k = 0; k < use->req(); k++) {
1121 Node* n = use->in(k);
1122 if (def == n) {
1123 if (!is_vector_use(use, k)) {
1124 return false;
1125 }
1126 }
1127 }
1128 }
1129 }
1130 }
1131 return true;
1132 }
1133
1134 //------------------------------schedule---------------------------
1135 // Adjust the memory graph for the packed operations
1136 void SuperWord::schedule() {
1137
1138 // Co-locate in the memory graph the members of each memory pack
1139 for (int i = 0; i < _packset.length(); i++) {
1140 co_locate_pack(_packset.at(i));
1141 }
1142 }
1143
1144 //-------------------------------remove_and_insert-------------------
1145 // Remove "current" from its current position in the memory graph and insert
1146 // it after the appropriate insertion point (lip or uip).
1147 void SuperWord::remove_and_insert(MemNode *current, MemNode *prev, MemNode *lip,
1148 Node *uip, Unique_Node_List &sched_before) {
1149 Node* my_mem = current->in(MemNode::Memory);
1150 bool sched_up = sched_before.member(current);
1151
1152 // remove current_store from its current position in the memmory graph
1153 for (DUIterator i = current->outs(); current->has_out(i); i++) {
1154 Node* use = current->out(i);
1155 if (use->is_Mem()) {
1156 assert(use->in(MemNode::Memory) == current, "must be");
1157 if (use == prev) { // connect prev to my_mem
1158 _igvn.replace_input_of(use, MemNode::Memory, my_mem);
1159 --i; //deleted this edge; rescan position
1160 } else if (sched_before.member(use)) {
1161 if (!sched_up) { // Will be moved together with current
1162 _igvn.replace_input_of(use, MemNode::Memory, uip);
1163 --i; //deleted this edge; rescan position
1164 }
1165 } else {
1166 if (sched_up) { // Will be moved together with current
1167 _igvn.replace_input_of(use, MemNode::Memory, lip);
1168 --i; //deleted this edge; rescan position
1169 }
1170 }
1171 }
1172 }
1173
1174 Node *insert_pt = sched_up ? uip : lip;
1175
1176 // all uses of insert_pt's memory state should use current's instead
1177 for (DUIterator i = insert_pt->outs(); insert_pt->has_out(i); i++) {
1178 Node* use = insert_pt->out(i);
1179 if (use->is_Mem()) {
1180 assert(use->in(MemNode::Memory) == insert_pt, "must be");
1181 _igvn.replace_input_of(use, MemNode::Memory, current);
1182 --i; //deleted this edge; rescan position
1183 } else if (!sched_up && use->is_Phi() && use->bottom_type() == Type::MEMORY) {
1184 uint pos; //lip (lower insert point) must be the last one in the memory slice
1185 for (pos=1; pos < use->req(); pos++) {
1186 if (use->in(pos) == insert_pt) break;
1187 }
1188 _igvn.replace_input_of(use, pos, current);
1189 --i;
1190 }
1191 }
1192
1193 //connect current to insert_pt
1194 _igvn.replace_input_of(current, MemNode::Memory, insert_pt);
1195 }
1196
1197 //------------------------------co_locate_pack----------------------------------
1198 // To schedule a store pack, we need to move any sandwiched memory ops either before
1199 // or after the pack, based upon dependence information:
1200 // (1) If any store in the pack depends on the sandwiched memory op, the
1201 // sandwiched memory op must be scheduled BEFORE the pack;
1202 // (2) If a sandwiched memory op depends on any store in the pack, the
1203 // sandwiched memory op must be scheduled AFTER the pack;
1204 // (3) If a sandwiched memory op (say, memA) depends on another sandwiched
1205 // memory op (say memB), memB must be scheduled before memA. So, if memA is
1206 // scheduled before the pack, memB must also be scheduled before the pack;
1207 // (4) If there is no dependence restriction for a sandwiched memory op, we simply
1208 // schedule this store AFTER the pack
1209 // (5) We know there is no dependence cycle, so there in no other case;
1210 // (6) Finally, all memory ops in another single pack should be moved in the same direction.
1211 //
1212 // To schedule a load pack, we use the memory state of either the first or the last load in
1213 // the pack, based on the dependence constraint.
1214 void SuperWord::co_locate_pack(Node_List* pk) {
1215 if (pk->at(0)->is_Store()) {
1216 MemNode* first = executed_first(pk)->as_Mem();
1217 MemNode* last = executed_last(pk)->as_Mem();
1218 Unique_Node_List schedule_before_pack;
1219 Unique_Node_List memops;
1220
1221 MemNode* current = last->in(MemNode::Memory)->as_Mem();
1222 MemNode* previous = last;
1223 while (true) {
1224 assert(in_bb(current), "stay in block");
1225 memops.push(previous);
1226 for (DUIterator i = current->outs(); current->has_out(i); i++) {
1227 Node* use = current->out(i);
1228 if (use->is_Mem() && use != previous)
1229 memops.push(use);
1230 }
1231 if (current == first) break;
1232 previous = current;
1233 current = current->in(MemNode::Memory)->as_Mem();
1234 }
1235
1236 // determine which memory operations should be scheduled before the pack
1237 for (uint i = 1; i < memops.size(); i++) {
1238 Node *s1 = memops.at(i);
1239 if (!in_pack(s1, pk) && !schedule_before_pack.member(s1)) {
1240 for (uint j = 0; j< i; j++) {
1241 Node *s2 = memops.at(j);
1242 if (!independent(s1, s2)) {
1243 if (in_pack(s2, pk) || schedule_before_pack.member(s2)) {
1244 schedule_before_pack.push(s1); // s1 must be scheduled before
1245 Node_List* mem_pk = my_pack(s1);
1246 if (mem_pk != NULL) {
1247 for (uint ii = 0; ii < mem_pk->size(); ii++) {
1248 Node* s = mem_pk->at(ii); // follow partner
1249 if (memops.member(s) && !schedule_before_pack.member(s))
1250 schedule_before_pack.push(s);
1251 }
1252 }
1253 break;
1254 }
1255 }
1256 }
1257 }
1258 }
1259
1260 Node* upper_insert_pt = first->in(MemNode::Memory);
1261 // Following code moves loads connected to upper_insert_pt below aliased stores.
1262 // Collect such loads here and reconnect them back to upper_insert_pt later.
1263 memops.clear();
1264 for (DUIterator i = upper_insert_pt->outs(); upper_insert_pt->has_out(i); i++) {
1265 Node* use = upper_insert_pt->out(i);
1266 if (!use->is_Store())
1267 memops.push(use);
1268 }
1269
1270 MemNode* lower_insert_pt = last;
1271 previous = last; //previous store in pk
1272 current = last->in(MemNode::Memory)->as_Mem();
1273
1274 // start scheduling from "last" to "first"
1275 while (true) {
1276 assert(in_bb(current), "stay in block");
1277 assert(in_pack(previous, pk), "previous stays in pack");
1278 Node* my_mem = current->in(MemNode::Memory);
1279
1280 if (in_pack(current, pk)) {
1281 // Forward users of my memory state (except "previous) to my input memory state
1282 for (DUIterator i = current->outs(); current->has_out(i); i++) {
1283 Node* use = current->out(i);
1284 if (use->is_Mem() && use != previous) {
1285 assert(use->in(MemNode::Memory) == current, "must be");
1286 if (schedule_before_pack.member(use)) {
1287 _igvn.replace_input_of(use, MemNode::Memory, upper_insert_pt);
1288 } else {
1289 _igvn.replace_input_of(use, MemNode::Memory, lower_insert_pt);
1290 }
1291 --i; // deleted this edge; rescan position
1292 }
1293 }
1294 previous = current;
1295 } else { // !in_pack(current, pk) ==> a sandwiched store
1296 remove_and_insert(current, previous, lower_insert_pt, upper_insert_pt, schedule_before_pack);
1297 }
1298
1299 if (current == first) break;
1300 current = my_mem->as_Mem();
1301 } // end while
1302
1303 // Reconnect loads back to upper_insert_pt.
1304 for (uint i = 0; i < memops.size(); i++) {
1305 Node *ld = memops.at(i);
1306 if (ld->in(MemNode::Memory) != upper_insert_pt) {
1307 _igvn.replace_input_of(ld, MemNode::Memory, upper_insert_pt);
1308 }
1309 }
1310 } else if (pk->at(0)->is_Load()) { //load
1311 // all loads in the pack should have the same memory state. By default,
1312 // we use the memory state of the last load. However, if any load could
1313 // not be moved down due to the dependence constraint, we use the memory
1314 // state of the first load.
1315 Node* last_mem = executed_last(pk)->in(MemNode::Memory);
1316 Node* first_mem = executed_first(pk)->in(MemNode::Memory);
1317 bool schedule_last = true;
1318 for (uint i = 0; i < pk->size(); i++) {
1319 Node* ld = pk->at(i);
1320 for (Node* current = last_mem; current != ld->in(MemNode::Memory);
1321 current=current->in(MemNode::Memory)) {
1322 assert(current != first_mem, "corrupted memory graph");
1323 if(current->is_Mem() && !independent(current, ld)){
1324 schedule_last = false; // a later store depends on this load
1325 break;
1326 }
1327 }
1328 }
1329
1330 Node* mem_input = schedule_last ? last_mem : first_mem;
1331 _igvn.hash_delete(mem_input);
1332 // Give each load the same memory state
1333 for (uint i = 0; i < pk->size(); i++) {
1334 LoadNode* ld = pk->at(i)->as_Load();
1335 _igvn.replace_input_of(ld, MemNode::Memory, mem_input);
1336 }
1337 }
1338 }
1339
1340 //------------------------------output---------------------------
1341 // Convert packs into vector node operations
1342 void SuperWord::output() {
1343 if (_packset.length() == 0) return;
1344
1345 #ifndef PRODUCT
1346 if (TraceLoopOpts) {
1347 tty->print("SuperWord ");
1348 lpt()->dump_head();
1349 }
1350 #endif
1351
1352 // MUST ENSURE main loop's initial value is properly aligned:
1353 // (iv_initial_value + min_iv_offset) % vector_width_in_bytes() == 0
1354
1355 align_initial_loop_index(align_to_ref());
1356
1357 // Insert extract (unpack) operations for scalar uses
1358 for (int i = 0; i < _packset.length(); i++) {
1359 insert_extracts(_packset.at(i));
1360 }
1361
1362 Compile* C = _phase->C;
1363 uint max_vlen_in_bytes = 0;
1364 for (int i = 0; i < _block.length(); i++) {
1365 Node* n = _block.at(i);
1366 Node_List* p = my_pack(n);
1367 if (p && n == executed_last(p)) {
1368 uint vlen = p->size();
1369 uint vlen_in_bytes = 0;
1370 Node* vn = NULL;
1371 Node* low_adr = p->at(0);
1372 Node* first = executed_first(p);
1373 int opc = n->Opcode();
1374 if (n->is_Load()) {
1375 Node* ctl = n->in(MemNode::Control);
1376 Node* mem = first->in(MemNode::Memory);
1377 Node* adr = low_adr->in(MemNode::Address);
1378 const TypePtr* atyp = n->adr_type();
1379 vn = LoadVectorNode::make(C, opc, ctl, mem, adr, atyp, vlen, velt_basic_type(n));
1380 vlen_in_bytes = vn->as_LoadVector()->memory_size();
1381 } else if (n->is_Store()) {
1382 // Promote value to be stored to vector
1383 Node* val = vector_opd(p, MemNode::ValueIn);
1384 Node* ctl = n->in(MemNode::Control);
1385 Node* mem = first->in(MemNode::Memory);
1386 Node* adr = low_adr->in(MemNode::Address);
1387 const TypePtr* atyp = n->adr_type();
1388 vn = StoreVectorNode::make(C, opc, ctl, mem, adr, atyp, val, vlen);
1389 vlen_in_bytes = vn->as_StoreVector()->memory_size();
1390 } else if (n->req() == 3) {
1391 // Promote operands to vector
1392 Node* in1 = vector_opd(p, 1);
1393 Node* in2 = vector_opd(p, 2);
1394 if (VectorNode::is_invariant_vector(in1) && (n->is_Add() || n->is_Mul())) {
1395 // Move invariant vector input into second position to avoid register spilling.
1396 Node* tmp = in1;
1397 in1 = in2;
1398 in2 = tmp;
1399 }
1400 vn = VectorNode::make(C, opc, in1, in2, vlen, velt_basic_type(n));
1401 vlen_in_bytes = vn->as_Vector()->length_in_bytes();
1402 } else {
1403 ShouldNotReachHere();
1404 }
1405 assert(vn != NULL, "sanity");
1406 _igvn.register_new_node_with_optimizer(vn);
1407 _phase->set_ctrl(vn, _phase->get_ctrl(p->at(0)));
1408 for (uint j = 0; j < p->size(); j++) {
1409 Node* pm = p->at(j);
1410 _igvn.replace_node(pm, vn);
1411 }
1412 _igvn._worklist.push(vn);
1413
1414 if (vlen_in_bytes > max_vlen_in_bytes) {
1415 max_vlen_in_bytes = vlen_in_bytes;
1416 }
1417 #ifdef ASSERT
1418 if (TraceNewVectors) {
1419 tty->print("new Vector node: ");
1420 vn->dump();
1421 }
1422 #endif
1423 }
1424 }
1425 C->set_max_vector_size(max_vlen_in_bytes);
1426 }
1427
1428 //------------------------------vector_opd---------------------------
1429 // Create a vector operand for the nodes in pack p for operand: in(opd_idx)
1430 Node* SuperWord::vector_opd(Node_List* p, int opd_idx) {
1431 Node* p0 = p->at(0);
1432 uint vlen = p->size();
1433 Node* opd = p0->in(opd_idx);
1434
1435 if (same_inputs(p, opd_idx)) {
1436 if (opd->is_Vector() || opd->is_LoadVector()) {
1437 assert(((opd_idx != 2) || !VectorNode::is_shift(p0)), "shift's count can't be vector");
1438 return opd; // input is matching vector
1439 }
1440 if ((opd_idx == 2) && VectorNode::is_shift(p0)) {
1441 Compile* C = _phase->C;
1442 Node* cnt = opd;
1443 // Vector instructions do not mask shift count, do it here.
1444 juint mask = (p0->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
1445 const TypeInt* t = opd->find_int_type();
1446 if (t != NULL && t->is_con()) {
1447 juint shift = t->get_con();
1448 if (shift > mask) { // Unsigned cmp
1449 cnt = ConNode::make(C, TypeInt::make(shift & mask));
1450 }
1451 } else {
1452 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
1453 cnt = ConNode::make(C, TypeInt::make(mask));
1454 _igvn.register_new_node_with_optimizer(cnt);
1455 cnt = new (C) AndINode(opd, cnt);
1456 _igvn.register_new_node_with_optimizer(cnt);
1457 _phase->set_ctrl(cnt, _phase->get_ctrl(opd));
1458 }
1459 assert(opd->bottom_type()->isa_int(), "int type only");
1460 // Move non constant shift count into vector register.
1461 cnt = VectorNode::shift_count(C, p0, cnt, vlen, velt_basic_type(p0));
1462 }
1463 if (cnt != opd) {
1464 _igvn.register_new_node_with_optimizer(cnt);
1465 _phase->set_ctrl(cnt, _phase->get_ctrl(opd));
1466 }
1467 return cnt;
1468 }
1469 assert(!opd->is_StoreVector(), "such vector is not expected here");
1470 // Convert scalar input to vector with the same number of elements as
1471 // p0's vector. Use p0's type because size of operand's container in
1472 // vector should match p0's size regardless operand's size.
1473 const Type* p0_t = velt_type(p0);
1474 VectorNode* vn = VectorNode::scalar2vector(_phase->C, opd, vlen, p0_t);
1475
1476 _igvn.register_new_node_with_optimizer(vn);
1477 _phase->set_ctrl(vn, _phase->get_ctrl(opd));
1478 #ifdef ASSERT
1479 if (TraceNewVectors) {
1480 tty->print("new Vector node: ");
1481 vn->dump();
1482 }
1483 #endif
1484 return vn;
1485 }
1486
1487 // Insert pack operation
1488 BasicType bt = velt_basic_type(p0);
1489 PackNode* pk = PackNode::make(_phase->C, opd, vlen, bt);
1490 DEBUG_ONLY( const BasicType opd_bt = opd->bottom_type()->basic_type(); )
1491
1492 for (uint i = 1; i < vlen; i++) {
1493 Node* pi = p->at(i);
1494 Node* in = pi->in(opd_idx);
1495 assert(my_pack(in) == NULL, "Should already have been unpacked");
1496 assert(opd_bt == in->bottom_type()->basic_type(), "all same type");
1497 pk->add_opd(in);
1498 }
1499 _igvn.register_new_node_with_optimizer(pk);
1500 _phase->set_ctrl(pk, _phase->get_ctrl(opd));
1501 #ifdef ASSERT
1502 if (TraceNewVectors) {
1503 tty->print("new Vector node: ");
1504 pk->dump();
1505 }
1506 #endif
1507 return pk;
1508 }
1509
1510 //------------------------------insert_extracts---------------------------
1511 // If a use of pack p is not a vector use, then replace the
1512 // use with an extract operation.
1513 void SuperWord::insert_extracts(Node_List* p) {
1514 if (p->at(0)->is_Store()) return;
1515 assert(_n_idx_list.is_empty(), "empty (node,index) list");
1516
1517 // Inspect each use of each pack member. For each use that is
1518 // not a vector use, replace the use with an extract operation.
1519
1520 for (uint i = 0; i < p->size(); i++) {
1521 Node* def = p->at(i);
1522 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
1523 Node* use = def->fast_out(j);
1524 for (uint k = 0; k < use->req(); k++) {
1525 Node* n = use->in(k);
1526 if (def == n) {
1527 if (!is_vector_use(use, k)) {
1528 _n_idx_list.push(use, k);
1529 }
1530 }
1531 }
1532 }
1533 }
1534
1535 while (_n_idx_list.is_nonempty()) {
1536 Node* use = _n_idx_list.node();
1537 int idx = _n_idx_list.index();
1538 _n_idx_list.pop();
1539 Node* def = use->in(idx);
1540
1541 // Insert extract operation
1542 _igvn.hash_delete(def);
1543 int def_pos = alignment(def) / data_size(def);
1544
1545 Node* ex = ExtractNode::make(_phase->C, def, def_pos, velt_basic_type(def));
1546 _igvn.register_new_node_with_optimizer(ex);
1547 _phase->set_ctrl(ex, _phase->get_ctrl(def));
1548 _igvn.replace_input_of(use, idx, ex);
1549 _igvn._worklist.push(def);
1550
1551 bb_insert_after(ex, bb_idx(def));
1552 set_velt_type(ex, velt_type(def));
1553 }
1554 }
1555
1556 //------------------------------is_vector_use---------------------------
1557 // Is use->in(u_idx) a vector use?
1558 bool SuperWord::is_vector_use(Node* use, int u_idx) {
1559 Node_List* u_pk = my_pack(use);
1560 if (u_pk == NULL) return false;
1561 Node* def = use->in(u_idx);
1562 Node_List* d_pk = my_pack(def);
1563 if (d_pk == NULL) {
1564 // check for scalar promotion
1565 Node* n = u_pk->at(0)->in(u_idx);
1566 for (uint i = 1; i < u_pk->size(); i++) {
1567 if (u_pk->at(i)->in(u_idx) != n) return false;
1568 }
1569 return true;
1570 }
1571 if (u_pk->size() != d_pk->size())
1572 return false;
1573 for (uint i = 0; i < u_pk->size(); i++) {
1574 Node* ui = u_pk->at(i);
1575 Node* di = d_pk->at(i);
1576 if (ui->in(u_idx) != di || alignment(ui) != alignment(di))
1577 return false;
1578 }
1579 return true;
1580 }
1581
1582 //------------------------------construct_bb---------------------------
1583 // Construct reverse postorder list of block members
1584 bool SuperWord::construct_bb() {
1585 Node* entry = bb();
1586
1587 assert(_stk.length() == 0, "stk is empty");
1588 assert(_block.length() == 0, "block is empty");
1589 assert(_data_entry.length() == 0, "data_entry is empty");
1590 assert(_mem_slice_head.length() == 0, "mem_slice_head is empty");
1591 assert(_mem_slice_tail.length() == 0, "mem_slice_tail is empty");
1592
1593 // Find non-control nodes with no inputs from within block,
1594 // create a temporary map from node _idx to bb_idx for use
1595 // by the visited and post_visited sets,
1596 // and count number of nodes in block.
1597 int bb_ct = 0;
1598 for (uint i = 0; i < lpt()->_body.size(); i++ ) {
1599 Node *n = lpt()->_body.at(i);
1600 set_bb_idx(n, i); // Create a temporary map
1601 if (in_bb(n)) {
1602 if (n->is_LoadStore() || n->is_MergeMem() ||
1603 (n->is_Proj() && !n->as_Proj()->is_CFG())) {
1604 // Bailout if the loop has LoadStore, MergeMem or data Proj
1605 // nodes. Superword optimization does not work with them.
1606 return false;
1607 }
1608 bb_ct++;
1609 if (!n->is_CFG()) {
1610 bool found = false;
1611 for (uint j = 0; j < n->req(); j++) {
1612 Node* def = n->in(j);
1613 if (def && in_bb(def)) {
1614 found = true;
1615 break;
1616 }
1617 }
1618 if (!found) {
1619 assert(n != entry, "can't be entry");
1620 _data_entry.push(n);
1621 }
1622 }
1623 }
1624 }
1625
1626 // Find memory slices (head and tail)
1627 for (DUIterator_Fast imax, i = lp()->fast_outs(imax); i < imax; i++) {
1628 Node *n = lp()->fast_out(i);
1629 if (in_bb(n) && (n->is_Phi() && n->bottom_type() == Type::MEMORY)) {
1630 Node* n_tail = n->in(LoopNode::LoopBackControl);
1631 if (n_tail != n->in(LoopNode::EntryControl)) {
1632 if (!n_tail->is_Mem()) {
1633 assert(n_tail->is_Mem(), err_msg_res("unexpected node for memory slice: %s", n_tail->Name()));
1634 return false; // Bailout
1635 }
1636 _mem_slice_head.push(n);
1637 _mem_slice_tail.push(n_tail);
1638 }
1639 }
1640 }
1641
1642 // Create an RPO list of nodes in block
1643
1644 visited_clear();
1645 post_visited_clear();
1646
1647 // Push all non-control nodes with no inputs from within block, then control entry
1648 for (int j = 0; j < _data_entry.length(); j++) {
1649 Node* n = _data_entry.at(j);
1650 visited_set(n);
1651 _stk.push(n);
1652 }
1653 visited_set(entry);
1654 _stk.push(entry);
1655
1656 // Do a depth first walk over out edges
1657 int rpo_idx = bb_ct - 1;
1658 int size;
1659 while ((size = _stk.length()) > 0) {
1660 Node* n = _stk.top(); // Leave node on stack
1661 if (!visited_test_set(n)) {
1662 // forward arc in graph
1663 } else if (!post_visited_test(n)) {
1664 // cross or back arc
1665 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1666 Node *use = n->fast_out(i);
1667 if (in_bb(use) && !visited_test(use) &&
1668 // Don't go around backedge
1669 (!use->is_Phi() || n == entry)) {
1670 _stk.push(use);
1671 }
1672 }
1673 if (_stk.length() == size) {
1674 // There were no additional uses, post visit node now
1675 _stk.pop(); // Remove node from stack
1676 assert(rpo_idx >= 0, "");
1677 _block.at_put_grow(rpo_idx, n);
1678 rpo_idx--;
1679 post_visited_set(n);
1680 assert(rpo_idx >= 0 || _stk.is_empty(), "");
1681 }
1682 } else {
1683 _stk.pop(); // Remove post-visited node from stack
1684 }
1685 }
1686
1687 // Create real map of block indices for nodes
1688 for (int j = 0; j < _block.length(); j++) {
1689 Node* n = _block.at(j);
1690 set_bb_idx(n, j);
1691 }
1692
1693 initialize_bb(); // Ensure extra info is allocated.
1694
1695 #ifndef PRODUCT
1696 if (TraceSuperWord) {
1697 print_bb();
1698 tty->print_cr("\ndata entry nodes: %s", _data_entry.length() > 0 ? "" : "NONE");
1699 for (int m = 0; m < _data_entry.length(); m++) {
1700 tty->print("%3d ", m);
1701 _data_entry.at(m)->dump();
1702 }
1703 tty->print_cr("\nmemory slices: %s", _mem_slice_head.length() > 0 ? "" : "NONE");
1704 for (int m = 0; m < _mem_slice_head.length(); m++) {
1705 tty->print("%3d ", m); _mem_slice_head.at(m)->dump();
1706 tty->print(" "); _mem_slice_tail.at(m)->dump();
1707 }
1708 }
1709 #endif
1710 assert(rpo_idx == -1 && bb_ct == _block.length(), "all block members found");
1711 return (_mem_slice_head.length() > 0) || (_data_entry.length() > 0);
1712 }
1713
1714 //------------------------------initialize_bb---------------------------
1715 // Initialize per node info
1716 void SuperWord::initialize_bb() {
1717 Node* last = _block.at(_block.length() - 1);
1718 grow_node_info(bb_idx(last));
1719 }
1720
1721 //------------------------------bb_insert_after---------------------------
1722 // Insert n into block after pos
1723 void SuperWord::bb_insert_after(Node* n, int pos) {
1724 int n_pos = pos + 1;
1725 // Make room
1726 for (int i = _block.length() - 1; i >= n_pos; i--) {
1727 _block.at_put_grow(i+1, _block.at(i));
1728 }
1729 for (int j = _node_info.length() - 1; j >= n_pos; j--) {
1730 _node_info.at_put_grow(j+1, _node_info.at(j));
1731 }
1732 // Set value
1733 _block.at_put_grow(n_pos, n);
1734 _node_info.at_put_grow(n_pos, SWNodeInfo::initial);
1735 // Adjust map from node->_idx to _block index
1736 for (int i = n_pos; i < _block.length(); i++) {
1737 set_bb_idx(_block.at(i), i);
1738 }
1739 }
1740
1741 //------------------------------compute_max_depth---------------------------
1742 // Compute max depth for expressions from beginning of block
1743 // Use to prune search paths during test for independence.
1744 void SuperWord::compute_max_depth() {
1745 int ct = 0;
1746 bool again;
1747 do {
1748 again = false;
1749 for (int i = 0; i < _block.length(); i++) {
1750 Node* n = _block.at(i);
1751 if (!n->is_Phi()) {
1752 int d_orig = depth(n);
1753 int d_in = 0;
1754 for (DepPreds preds(n, _dg); !preds.done(); preds.next()) {
1755 Node* pred = preds.current();
1756 if (in_bb(pred)) {
1757 d_in = MAX2(d_in, depth(pred));
1758 }
1759 }
1760 if (d_in + 1 != d_orig) {
1761 set_depth(n, d_in + 1);
1762 again = true;
1763 }
1764 }
1765 }
1766 ct++;
1767 } while (again);
1768 #ifndef PRODUCT
1769 if (TraceSuperWord && Verbose)
1770 tty->print_cr("compute_max_depth iterated: %d times", ct);
1771 #endif
1772 }
1773
1774 //-------------------------compute_vector_element_type-----------------------
1775 // Compute necessary vector element type for expressions
1776 // This propagates backwards a narrower integer type when the
1777 // upper bits of the value are not needed.
1778 // Example: char a,b,c; a = b + c;
1779 // Normally the type of the add is integer, but for packed character
1780 // operations the type of the add needs to be char.
1781 void SuperWord::compute_vector_element_type() {
1782 #ifndef PRODUCT
1783 if (TraceSuperWord && Verbose)
1784 tty->print_cr("\ncompute_velt_type:");
1785 #endif
1786
1787 // Initial type
1788 for (int i = 0; i < _block.length(); i++) {
1789 Node* n = _block.at(i);
1790 set_velt_type(n, container_type(n));
1791 }
1792
1793 // Propagate integer narrowed type backwards through operations
1794 // that don't depend on higher order bits
1795 for (int i = _block.length() - 1; i >= 0; i--) {
1796 Node* n = _block.at(i);
1797 // Only integer types need be examined
1798 const Type* vtn = velt_type(n);
1799 if (vtn->basic_type() == T_INT) {
1800 uint start, end;
1801 VectorNode::vector_operands(n, &start, &end);
1802
1803 for (uint j = start; j < end; j++) {
1804 Node* in = n->in(j);
1805 // Don't propagate through a memory
1806 if (!in->is_Mem() && in_bb(in) && velt_type(in)->basic_type() == T_INT &&
1807 data_size(n) < data_size(in)) {
1808 bool same_type = true;
1809 for (DUIterator_Fast kmax, k = in->fast_outs(kmax); k < kmax; k++) {
1810 Node *use = in->fast_out(k);
1811 if (!in_bb(use) || !same_velt_type(use, n)) {
1812 same_type = false;
1813 break;
1814 }
1815 }
1816 if (same_type) {
1817 // For right shifts of small integer types (bool, byte, char, short)
1818 // we need precise information about sign-ness. Only Load nodes have
1819 // this information because Store nodes are the same for signed and
1820 // unsigned values. And any arithmetic operation after a load may
1821 // expand a value to signed Int so such right shifts can't be used
1822 // because vector elements do not have upper bits of Int.
1823 const Type* vt = vtn;
1824 if (VectorNode::is_shift(in)) {
1825 Node* load = in->in(1);
1826 if (load->is_Load() && in_bb(load) && (velt_type(load)->basic_type() == T_INT)) {
1827 vt = velt_type(load);
1828 } else if (in->Opcode() != Op_LShiftI) {
1829 // Widen type to Int to avoid creation of right shift vector
1830 // (align + data_size(s1) check in stmts_can_pack() will fail).
1831 // Note, left shifts work regardless type.
1832 vt = TypeInt::INT;
1833 }
1834 }
1835 set_velt_type(in, vt);
1836 }
1837 }
1838 }
1839 }
1840 }
1841 #ifndef PRODUCT
1842 if (TraceSuperWord && Verbose) {
1843 for (int i = 0; i < _block.length(); i++) {
1844 Node* n = _block.at(i);
1845 velt_type(n)->dump();
1846 tty->print("\t");
1847 n->dump();
1848 }
1849 }
1850 #endif
1851 }
1852
1853 //------------------------------memory_alignment---------------------------
1854 // Alignment within a vector memory reference
1855 int SuperWord::memory_alignment(MemNode* s, int iv_adjust) {
1856 SWPointer p(s, this);
1857 if (!p.valid()) {
1858 return bottom_align;
1859 }
1860 int vw = vector_width_in_bytes(s);
1861 if (vw < 2) {
1862 return bottom_align; // No vectors for this type
1863 }
1864 int offset = p.offset_in_bytes();
1865 offset += iv_adjust*p.memory_size();
1866 int off_rem = offset % vw;
1867 int off_mod = off_rem >= 0 ? off_rem : off_rem + vw;
1868 return off_mod;
1869 }
1870
1871 //---------------------------container_type---------------------------
1872 // Smallest type containing range of values
1873 const Type* SuperWord::container_type(Node* n) {
1874 if (n->is_Mem()) {
1875 BasicType bt = n->as_Mem()->memory_type();
1876 if (n->is_Store() && (bt == T_CHAR)) {
1877 // Use T_SHORT type instead of T_CHAR for stored values because any
1878 // preceding arithmetic operation extends values to signed Int.
1879 bt = T_SHORT;
1880 }
1881 if (n->Opcode() == Op_LoadUB) {
1882 // Adjust type for unsigned byte loads, it is important for right shifts.
1883 // T_BOOLEAN is used because there is no basic type representing type
1884 // TypeInt::UBYTE. Use of T_BOOLEAN for vectors is fine because only
1885 // size (one byte) and sign is important.
1886 bt = T_BOOLEAN;
1887 }
1888 return Type::get_const_basic_type(bt);
1889 }
1890 const Type* t = _igvn.type(n);
1891 if (t->basic_type() == T_INT) {
1892 // A narrow type of arithmetic operations will be determined by
1893 // propagating the type of memory operations.
1894 return TypeInt::INT;
1895 }
1896 return t;
1897 }
1898
1899 bool SuperWord::same_velt_type(Node* n1, Node* n2) {
1900 const Type* vt1 = velt_type(n1);
1901 const Type* vt2 = velt_type(n2);
1902 if (vt1->basic_type() == T_INT && vt2->basic_type() == T_INT) {
1903 // Compare vectors element sizes for integer types.
1904 return data_size(n1) == data_size(n2);
1905 }
1906 return vt1 == vt2;
1907 }
1908
1909 //------------------------------in_packset---------------------------
1910 // Are s1 and s2 in a pack pair and ordered as s1,s2?
1911 bool SuperWord::in_packset(Node* s1, Node* s2) {
1912 for (int i = 0; i < _packset.length(); i++) {
1913 Node_List* p = _packset.at(i);
1914 assert(p->size() == 2, "must be");
1915 if (p->at(0) == s1 && p->at(p->size()-1) == s2) {
1916 return true;
1917 }
1918 }
1919 return false;
1920 }
1921
1922 //------------------------------in_pack---------------------------
1923 // Is s in pack p?
1924 Node_List* SuperWord::in_pack(Node* s, Node_List* p) {
1925 for (uint i = 0; i < p->size(); i++) {
1926 if (p->at(i) == s) {
1927 return p;
1928 }
1929 }
1930 return NULL;
1931 }
1932
1933 //------------------------------remove_pack_at---------------------------
1934 // Remove the pack at position pos in the packset
1935 void SuperWord::remove_pack_at(int pos) {
1936 Node_List* p = _packset.at(pos);
1937 for (uint i = 0; i < p->size(); i++) {
1938 Node* s = p->at(i);
1939 set_my_pack(s, NULL);
1940 }
1941 _packset.remove_at(pos);
1942 }
1943
1944 //------------------------------executed_first---------------------------
1945 // Return the node executed first in pack p. Uses the RPO block list
1946 // to determine order.
1947 Node* SuperWord::executed_first(Node_List* p) {
1948 Node* n = p->at(0);
1949 int n_rpo = bb_idx(n);
1950 for (uint i = 1; i < p->size(); i++) {
1951 Node* s = p->at(i);
1952 int s_rpo = bb_idx(s);
1953 if (s_rpo < n_rpo) {
1954 n = s;
1955 n_rpo = s_rpo;
1956 }
1957 }
1958 return n;
1959 }
1960
1961 //------------------------------executed_last---------------------------
1962 // Return the node executed last in pack p.
1963 Node* SuperWord::executed_last(Node_List* p) {
1964 Node* n = p->at(0);
1965 int n_rpo = bb_idx(n);
1966 for (uint i = 1; i < p->size(); i++) {
1967 Node* s = p->at(i);
1968 int s_rpo = bb_idx(s);
1969 if (s_rpo > n_rpo) {
1970 n = s;
1971 n_rpo = s_rpo;
1972 }
1973 }
1974 return n;
1975 }
1976
1977 //----------------------------align_initial_loop_index---------------------------
1978 // Adjust pre-loop limit so that in main loop, a load/store reference
1979 // to align_to_ref will be a position zero in the vector.
1980 // (iv + k) mod vector_align == 0
1981 void SuperWord::align_initial_loop_index(MemNode* align_to_ref) {
1982 CountedLoopNode *main_head = lp()->as_CountedLoop();
1983 assert(main_head->is_main_loop(), "");
1984 CountedLoopEndNode* pre_end = get_pre_loop_end(main_head);
1985 assert(pre_end != NULL, "we must have a correct pre-loop");
1986 Node *pre_opaq1 = pre_end->limit();
1987 assert(pre_opaq1->Opcode() == Op_Opaque1, "");
1988 Opaque1Node *pre_opaq = (Opaque1Node*)pre_opaq1;
1989 Node *lim0 = pre_opaq->in(1);
1990
1991 // Where we put new limit calculations
1992 Node *pre_ctrl = pre_end->loopnode()->in(LoopNode::EntryControl);
1993
1994 // Ensure the original loop limit is available from the
1995 // pre-loop Opaque1 node.
1996 Node *orig_limit = pre_opaq->original_loop_limit();
1997 assert(orig_limit != NULL && _igvn.type(orig_limit) != Type::TOP, "");
1998
1999 SWPointer align_to_ref_p(align_to_ref, this);
2000 assert(align_to_ref_p.valid(), "sanity");
2001
2002 // Given:
2003 // lim0 == original pre loop limit
2004 // V == v_align (power of 2)
2005 // invar == extra invariant piece of the address expression
2006 // e == offset [ +/- invar ]
2007 //
2008 // When reassociating expressions involving '%' the basic rules are:
2009 // (a - b) % k == 0 => a % k == b % k
2010 // and:
2011 // (a + b) % k == 0 => a % k == (k - b) % k
2012 //
2013 // For stride > 0 && scale > 0,
2014 // Derive the new pre-loop limit "lim" such that the two constraints:
2015 // (1) lim = lim0 + N (where N is some positive integer < V)
2016 // (2) (e + lim) % V == 0
2017 // are true.
2018 //
2019 // Substituting (1) into (2),
2020 // (e + lim0 + N) % V == 0
2021 // solve for N:
2022 // N = (V - (e + lim0)) % V
2023 // substitute back into (1), so that new limit
2024 // lim = lim0 + (V - (e + lim0)) % V
2025 //
2026 // For stride > 0 && scale < 0
2027 // Constraints:
2028 // lim = lim0 + N
2029 // (e - lim) % V == 0
2030 // Solving for lim:
2031 // (e - lim0 - N) % V == 0
2032 // N = (e - lim0) % V
2033 // lim = lim0 + (e - lim0) % V
2034 //
2035 // For stride < 0 && scale > 0
2036 // Constraints:
2037 // lim = lim0 - N
2038 // (e + lim) % V == 0
2039 // Solving for lim:
2040 // (e + lim0 - N) % V == 0
2041 // N = (e + lim0) % V
2042 // lim = lim0 - (e + lim0) % V
2043 //
2044 // For stride < 0 && scale < 0
2045 // Constraints:
2046 // lim = lim0 - N
2047 // (e - lim) % V == 0
2048 // Solving for lim:
2049 // (e - lim0 + N) % V == 0
2050 // N = (V - (e - lim0)) % V
2051 // lim = lim0 - (V - (e - lim0)) % V
2052
2053 int vw = vector_width_in_bytes(align_to_ref);
2054 int stride = iv_stride();
2055 int scale = align_to_ref_p.scale_in_bytes();
2056 int elt_size = align_to_ref_p.memory_size();
2057 int v_align = vw / elt_size;
2058 assert(v_align > 1, "sanity");
2059 int offset = align_to_ref_p.offset_in_bytes() / elt_size;
2060 Node *offsn = _igvn.intcon(offset);
2061
2062 Node *e = offsn;
2063 if (align_to_ref_p.invar() != NULL) {
2064 // incorporate any extra invariant piece producing (offset +/- invar) >>> log2(elt)
2065 Node* log2_elt = _igvn.intcon(exact_log2(elt_size));
2066 Node* aref = new (_phase->C) URShiftINode(align_to_ref_p.invar(), log2_elt);
2067 _igvn.register_new_node_with_optimizer(aref);
2068 _phase->set_ctrl(aref, pre_ctrl);
2069 if (align_to_ref_p.negate_invar()) {
2070 e = new (_phase->C) SubINode(e, aref);
2071 } else {
2072 e = new (_phase->C) AddINode(e, aref);
2073 }
2074 _igvn.register_new_node_with_optimizer(e);
2075 _phase->set_ctrl(e, pre_ctrl);
2076 }
2077 if (vw > ObjectAlignmentInBytes) {
2078 // incorporate base e +/- base && Mask >>> log2(elt)
2079 Node* xbase = new(_phase->C) CastP2XNode(NULL, align_to_ref_p.base());
2080 _igvn.register_new_node_with_optimizer(xbase);
2081 #ifdef _LP64
2082 xbase = new (_phase->C) ConvL2INode(xbase);
2083 _igvn.register_new_node_with_optimizer(xbase);
2084 #endif
2085 Node* mask = _igvn.intcon(vw-1);
2086 Node* masked_xbase = new (_phase->C) AndINode(xbase, mask);
2087 _igvn.register_new_node_with_optimizer(masked_xbase);
2088 Node* log2_elt = _igvn.intcon(exact_log2(elt_size));
2089 Node* bref = new (_phase->C) URShiftINode(masked_xbase, log2_elt);
2090 _igvn.register_new_node_with_optimizer(bref);
2091 _phase->set_ctrl(bref, pre_ctrl);
2092 e = new (_phase->C) AddINode(e, bref);
2093 _igvn.register_new_node_with_optimizer(e);
2094 _phase->set_ctrl(e, pre_ctrl);
2095 }
2096
2097 // compute e +/- lim0
2098 if (scale < 0) {
2099 e = new (_phase->C) SubINode(e, lim0);
2100 } else {
2101 e = new (_phase->C) AddINode(e, lim0);
2102 }
2103 _igvn.register_new_node_with_optimizer(e);
2104 _phase->set_ctrl(e, pre_ctrl);
2105
2106 if (stride * scale > 0) {
2107 // compute V - (e +/- lim0)
2108 Node* va = _igvn.intcon(v_align);
2109 e = new (_phase->C) SubINode(va, e);
2110 _igvn.register_new_node_with_optimizer(e);
2111 _phase->set_ctrl(e, pre_ctrl);
2112 }
2113 // compute N = (exp) % V
2114 Node* va_msk = _igvn.intcon(v_align - 1);
2115 Node* N = new (_phase->C) AndINode(e, va_msk);
2116 _igvn.register_new_node_with_optimizer(N);
2117 _phase->set_ctrl(N, pre_ctrl);
2118
2119 // substitute back into (1), so that new limit
2120 // lim = lim0 + N
2121 Node* lim;
2122 if (stride < 0) {
2123 lim = new (_phase->C) SubINode(lim0, N);
2124 } else {
2125 lim = new (_phase->C) AddINode(lim0, N);
2126 }
2127 _igvn.register_new_node_with_optimizer(lim);
2128 _phase->set_ctrl(lim, pre_ctrl);
2129 Node* constrained =
2130 (stride > 0) ? (Node*) new (_phase->C) MinINode(lim, orig_limit)
2131 : (Node*) new (_phase->C) MaxINode(lim, orig_limit);
2132 _igvn.register_new_node_with_optimizer(constrained);
2133 _phase->set_ctrl(constrained, pre_ctrl);
2134 _igvn.hash_delete(pre_opaq);
2135 pre_opaq->set_req(1, constrained);
2136 }
2137
2138 //----------------------------get_pre_loop_end---------------------------
2139 // Find pre loop end from main loop. Returns null if none.
2140 CountedLoopEndNode* SuperWord::get_pre_loop_end(CountedLoopNode *cl) {
2141 Node *ctrl = cl->in(LoopNode::EntryControl);
2142 if (!ctrl->is_IfTrue() && !ctrl->is_IfFalse()) return NULL;
2143 Node *iffm = ctrl->in(0);
2144 if (!iffm->is_If()) return NULL;
2145 Node *p_f = iffm->in(0);
2146 if (!p_f->is_IfFalse()) return NULL;
2147 if (!p_f->in(0)->is_CountedLoopEnd()) return NULL;
2148 CountedLoopEndNode *pre_end = p_f->in(0)->as_CountedLoopEnd();
2149 CountedLoopNode* loop_node = pre_end->loopnode();
2150 if (loop_node == NULL || !loop_node->is_pre_loop()) return NULL;
2151 return pre_end;
2152 }
2153
2154
2155 //------------------------------init---------------------------
2156 void SuperWord::init() {
2157 _dg.init();
2158 _packset.clear();
2159 _disjoint_ptrs.clear();
2160 _block.clear();
2161 _data_entry.clear();
2162 _mem_slice_head.clear();
2163 _mem_slice_tail.clear();
2164 _node_info.clear();
2165 _align_to_ref = NULL;
2166 _lpt = NULL;
2167 _lp = NULL;
2168 _bb = NULL;
2169 _iv = NULL;
2170 }
2171
2172 //------------------------------print_packset---------------------------
2173 void SuperWord::print_packset() {
2174 #ifndef PRODUCT
2175 tty->print_cr("packset");
2176 for (int i = 0; i < _packset.length(); i++) {
2177 tty->print_cr("Pack: %d", i);
2178 Node_List* p = _packset.at(i);
2179 print_pack(p);
2180 }
2181 #endif
2182 }
2183
2184 //------------------------------print_pack---------------------------
2185 void SuperWord::print_pack(Node_List* p) {
2186 for (uint i = 0; i < p->size(); i++) {
2187 print_stmt(p->at(i));
2188 }
2189 }
2190
2191 //------------------------------print_bb---------------------------
2192 void SuperWord::print_bb() {
2193 #ifndef PRODUCT
2194 tty->print_cr("\nBlock");
2195 for (int i = 0; i < _block.length(); i++) {
2196 Node* n = _block.at(i);
2197 tty->print("%d ", i);
2198 if (n) {
2199 n->dump();
2200 }
2201 }
2202 #endif
2203 }
2204
2205 //------------------------------print_stmt---------------------------
2206 void SuperWord::print_stmt(Node* s) {
2207 #ifndef PRODUCT
2208 tty->print(" align: %d \t", alignment(s));
2209 s->dump();
2210 #endif
2211 }
2212
2213 //------------------------------blank---------------------------
2214 char* SuperWord::blank(uint depth) {
2215 static char blanks[101];
2216 assert(depth < 101, "too deep");
2217 for (uint i = 0; i < depth; i++) blanks[i] = ' ';
2218 blanks[depth] = '\0';
2219 return blanks;
2220 }
2221
2222
2223 //==============================SWPointer===========================
2224
2225 //----------------------------SWPointer------------------------
2226 SWPointer::SWPointer(MemNode* mem, SuperWord* slp) :
2227 _mem(mem), _slp(slp), _base(NULL), _adr(NULL),
2228 _scale(0), _offset(0), _invar(NULL), _negate_invar(false) {
2229
2230 Node* adr = mem->in(MemNode::Address);
2231 if (!adr->is_AddP()) {
2232 assert(!valid(), "too complex");
2233 return;
2234 }
2235 // Match AddP(base, AddP(ptr, k*iv [+ invariant]), constant)
2236 Node* base = adr->in(AddPNode::Base);
2237 //unsafe reference could not be aligned appropriately without runtime checking
2238 if (base == NULL || base->bottom_type() == Type::TOP) {
2239 assert(!valid(), "unsafe access");
2240 return;
2241 }
2242 for (int i = 0; i < 3; i++) {
2243 if (!scaled_iv_plus_offset(adr->in(AddPNode::Offset))) {
2244 assert(!valid(), "too complex");
2245 return;
2246 }
2247 adr = adr->in(AddPNode::Address);
2248 if (base == adr || !adr->is_AddP()) {
2249 break; // stop looking at addp's
2250 }
2251 }
2252 _base = base;
2253 _adr = adr;
2254 assert(valid(), "Usable");
2255 }
2256
2257 // Following is used to create a temporary object during
2258 // the pattern match of an address expression.
2259 SWPointer::SWPointer(SWPointer* p) :
2260 _mem(p->_mem), _slp(p->_slp), _base(NULL), _adr(NULL),
2261 _scale(0), _offset(0), _invar(NULL), _negate_invar(false) {}
2262
2263 //------------------------scaled_iv_plus_offset--------------------
2264 // Match: k*iv + offset
2265 // where: k is a constant that maybe zero, and
2266 // offset is (k2 [+/- invariant]) where k2 maybe zero and invariant is optional
2267 bool SWPointer::scaled_iv_plus_offset(Node* n) {
2268 if (scaled_iv(n)) {
2269 return true;
2270 }
2271 if (offset_plus_k(n)) {
2272 return true;
2273 }
2274 int opc = n->Opcode();
2275 if (opc == Op_AddI) {
2276 if (scaled_iv(n->in(1)) && offset_plus_k(n->in(2))) {
2277 return true;
2278 }
2279 if (scaled_iv(n->in(2)) && offset_plus_k(n->in(1))) {
2280 return true;
2281 }
2282 } else if (opc == Op_SubI) {
2283 if (scaled_iv(n->in(1)) && offset_plus_k(n->in(2), true)) {
2284 return true;
2285 }
2286 if (scaled_iv(n->in(2)) && offset_plus_k(n->in(1))) {
2287 _scale *= -1;
2288 return true;
2289 }
2290 }
2291 return false;
2292 }
2293
2294 //----------------------------scaled_iv------------------------
2295 // Match: k*iv where k is a constant that's not zero
2296 bool SWPointer::scaled_iv(Node* n) {
2297 if (_scale != 0) {
2298 return false; // already found a scale
2299 }
2300 if (n == iv()) {
2301 _scale = 1;
2302 return true;
2303 }
2304 int opc = n->Opcode();
2305 if (opc == Op_MulI) {
2306 if (n->in(1) == iv() && n->in(2)->is_Con()) {
2307 _scale = n->in(2)->get_int();
2308 return true;
2309 } else if (n->in(2) == iv() && n->in(1)->is_Con()) {
2310 _scale = n->in(1)->get_int();
2311 return true;
2312 }
2313 } else if (opc == Op_LShiftI) {
2314 if (n->in(1) == iv() && n->in(2)->is_Con()) {
2315 _scale = 1 << n->in(2)->get_int();
2316 return true;
2317 }
2318 } else if (opc == Op_ConvI2L) {
2319 if (scaled_iv_plus_offset(n->in(1))) {
2320 return true;
2321 }
2322 } else if (opc == Op_LShiftL) {
2323 if (!has_iv() && _invar == NULL) {
2324 // Need to preserve the current _offset value, so
2325 // create a temporary object for this expression subtree.
2326 // Hacky, so should re-engineer the address pattern match.
2327 SWPointer tmp(this);
2328 if (tmp.scaled_iv_plus_offset(n->in(1))) {
2329 if (tmp._invar == NULL) {
2330 int mult = 1 << n->in(2)->get_int();
2331 _scale = tmp._scale * mult;
2332 _offset += tmp._offset * mult;
2333 return true;
2334 }
2335 }
2336 }
2337 }
2338 return false;
2339 }
2340
2341 //----------------------------offset_plus_k------------------------
2342 // Match: offset is (k [+/- invariant])
2343 // where k maybe zero and invariant is optional, but not both.
2344 bool SWPointer::offset_plus_k(Node* n, bool negate) {
2345 int opc = n->Opcode();
2346 if (opc == Op_ConI) {
2347 _offset += negate ? -(n->get_int()) : n->get_int();
2348 return true;
2349 } else if (opc == Op_ConL) {
2350 // Okay if value fits into an int
2351 const TypeLong* t = n->find_long_type();
2352 if (t->higher_equal(TypeLong::INT)) {
2353 jlong loff = n->get_long();
2354 jint off = (jint)loff;
2355 _offset += negate ? -off : loff;
2356 return true;
2357 }
2358 return false;
2359 }
2360 if (_invar != NULL) return false; // already have an invariant
2361 if (opc == Op_AddI) {
2362 if (n->in(2)->is_Con() && invariant(n->in(1))) {
2363 _negate_invar = negate;
2364 _invar = n->in(1);
2365 _offset += negate ? -(n->in(2)->get_int()) : n->in(2)->get_int();
2366 return true;
2367 } else if (n->in(1)->is_Con() && invariant(n->in(2))) {
2368 _offset += negate ? -(n->in(1)->get_int()) : n->in(1)->get_int();
2369 _negate_invar = negate;
2370 _invar = n->in(2);
2371 return true;
2372 }
2373 }
2374 if (opc == Op_SubI) {
2375 if (n->in(2)->is_Con() && invariant(n->in(1))) {
2376 _negate_invar = negate;
2377 _invar = n->in(1);
2378 _offset += !negate ? -(n->in(2)->get_int()) : n->in(2)->get_int();
2379 return true;
2380 } else if (n->in(1)->is_Con() && invariant(n->in(2))) {
2381 _offset += negate ? -(n->in(1)->get_int()) : n->in(1)->get_int();
2382 _negate_invar = !negate;
2383 _invar = n->in(2);
2384 return true;
2385 }
2386 }
2387 if (invariant(n)) {
2388 _negate_invar = negate;
2389 _invar = n;
2390 return true;
2391 }
2392 return false;
2393 }
2394
2395 //----------------------------print------------------------
2396 void SWPointer::print() {
2397 #ifndef PRODUCT
2398 tty->print("base: %d adr: %d scale: %d offset: %d invar: %c%d\n",
2399 _base != NULL ? _base->_idx : 0,
2400 _adr != NULL ? _adr->_idx : 0,
2401 _scale, _offset,
2402 _negate_invar?'-':'+',
2403 _invar != NULL ? _invar->_idx : 0);
2404 #endif
2405 }
2406
2407 // ========================= OrderedPair =====================
2408
2409 const OrderedPair OrderedPair::initial;
2410
2411 // ========================= SWNodeInfo =====================
2412
2413 const SWNodeInfo SWNodeInfo::initial;
2414
2415
2416 // ============================ DepGraph ===========================
2417
2418 //------------------------------make_node---------------------------
2419 // Make a new dependence graph node for an ideal node.
2420 DepMem* DepGraph::make_node(Node* node) {
2421 DepMem* m = new (_arena) DepMem(node);
2422 if (node != NULL) {
2423 assert(_map.at_grow(node->_idx) == NULL, "one init only");
2424 _map.at_put_grow(node->_idx, m);
2425 }
2426 return m;
2427 }
2428
2429 //------------------------------make_edge---------------------------
2430 // Make a new dependence graph edge from dpred -> dsucc
2431 DepEdge* DepGraph::make_edge(DepMem* dpred, DepMem* dsucc) {
2432 DepEdge* e = new (_arena) DepEdge(dpred, dsucc, dsucc->in_head(), dpred->out_head());
2433 dpred->set_out_head(e);
2434 dsucc->set_in_head(e);
2435 return e;
2436 }
2437
2438 // ========================== DepMem ========================
2439
2440 //------------------------------in_cnt---------------------------
2441 int DepMem::in_cnt() {
2442 int ct = 0;
2443 for (DepEdge* e = _in_head; e != NULL; e = e->next_in()) ct++;
2444 return ct;
2445 }
2446
2447 //------------------------------out_cnt---------------------------
2448 int DepMem::out_cnt() {
2449 int ct = 0;
2450 for (DepEdge* e = _out_head; e != NULL; e = e->next_out()) ct++;
2451 return ct;
2452 }
2453
2454 //------------------------------print-----------------------------
2455 void DepMem::print() {
2456 #ifndef PRODUCT
2457 tty->print(" DepNode %d (", _node->_idx);
2458 for (DepEdge* p = _in_head; p != NULL; p = p->next_in()) {
2459 Node* pred = p->pred()->node();
2460 tty->print(" %d", pred != NULL ? pred->_idx : 0);
2461 }
2462 tty->print(") [");
2463 for (DepEdge* s = _out_head; s != NULL; s = s->next_out()) {
2464 Node* succ = s->succ()->node();
2465 tty->print(" %d", succ != NULL ? succ->_idx : 0);
2466 }
2467 tty->print_cr(" ]");
2468 #endif
2469 }
2470
2471 // =========================== DepEdge =========================
2472
2473 //------------------------------DepPreds---------------------------
2474 void DepEdge::print() {
2475 #ifndef PRODUCT
2476 tty->print_cr("DepEdge: %d [ %d ]", _pred->node()->_idx, _succ->node()->_idx);
2477 #endif
2478 }
2479
2480 // =========================== DepPreds =========================
2481 // Iterator over predecessor edges in the dependence graph.
2482
2483 //------------------------------DepPreds---------------------------
2484 DepPreds::DepPreds(Node* n, DepGraph& dg) {
2485 _n = n;
2486 _done = false;
2487 if (_n->is_Store() || _n->is_Load()) {
2488 _next_idx = MemNode::Address;
2489 _end_idx = n->req();
2490 _dep_next = dg.dep(_n)->in_head();
2491 } else if (_n->is_Mem()) {
2492 _next_idx = 0;
2493 _end_idx = 0;
2494 _dep_next = dg.dep(_n)->in_head();
2495 } else {
2496 _next_idx = 1;
2497 _end_idx = _n->req();
2498 _dep_next = NULL;
2499 }
2500 next();
2501 }
2502
2503 //------------------------------next---------------------------
2504 void DepPreds::next() {
2505 if (_dep_next != NULL) {
2506 _current = _dep_next->pred()->node();
2507 _dep_next = _dep_next->next_in();
2508 } else if (_next_idx < _end_idx) {
2509 _current = _n->in(_next_idx++);
2510 } else {
2511 _done = true;
2512 }
2513 }
2514
2515 // =========================== DepSuccs =========================
2516 // Iterator over successor edges in the dependence graph.
2517
2518 //------------------------------DepSuccs---------------------------
2519 DepSuccs::DepSuccs(Node* n, DepGraph& dg) {
2520 _n = n;
2521 _done = false;
2522 if (_n->is_Load()) {
2523 _next_idx = 0;
2524 _end_idx = _n->outcnt();
2525 _dep_next = dg.dep(_n)->out_head();
2526 } else if (_n->is_Mem() || _n->is_Phi() && _n->bottom_type() == Type::MEMORY) {
2527 _next_idx = 0;
2528 _end_idx = 0;
2529 _dep_next = dg.dep(_n)->out_head();
2530 } else {
2531 _next_idx = 0;
2532 _end_idx = _n->outcnt();
2533 _dep_next = NULL;
2534 }
2535 next();
2536 }
2537
2538 //-------------------------------next---------------------------
2539 void DepSuccs::next() {
2540 if (_dep_next != NULL) {
2541 _current = _dep_next->succ()->node();
2542 _dep_next = _dep_next->next_out();
2543 } else if (_next_idx < _end_idx) {
2544 _current = _n->raw_out(_next_idx++);
2545 } else {
2546 _done = true;
2547 }
2548 }