1 /*
2 * Copyright (c) 2007, 2012, 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 construct_bb();
147
148 dependence_graph();
149
150 compute_max_depth();
151
152 compute_vector_element_type();
153
154 // Attempt vectorization
155
156 find_adjacent_refs();
157
158 extend_packlist();
159
160 combine_packs();
161
162 construct_my_pack_map();
163
164 filter_packs();
165
166 schedule();
167
168 output();
169 }
170
171 //------------------------------find_adjacent_refs---------------------------
172 // Find the adjacent memory references and create pack pairs for them.
173 // This is the initial set of packs that will then be extended by
174 // following use->def and def->use links. The align positions are
175 // assigned relative to the reference "align_to_ref"
176 void SuperWord::find_adjacent_refs() {
177 // Get list of memory operations
178 Node_List memops;
179 for (int i = 0; i < _block.length(); i++) {
180 Node* n = _block.at(i);
181 if (n->is_Mem() && !n->is_LoadStore() && in_bb(n) &&
182 is_java_primitive(n->as_Mem()->memory_type())) {
183 int align = memory_alignment(n->as_Mem(), 0);
184 if (align != bottom_align) {
185 memops.push(n);
186 }
187 }
188 }
189
190 Node_List align_to_refs;
191 int best_iv_adjustment = 0;
192 MemNode* best_align_to_mem_ref = NULL;
193
194 while (memops.size() != 0) {
195 // Find a memory reference to align to.
196 MemNode* mem_ref = find_align_to_ref(memops);
197 if (mem_ref == NULL) break;
198 align_to_refs.push(mem_ref);
199 int iv_adjustment = get_iv_adjustment(mem_ref);
200
201 if (best_align_to_mem_ref == NULL) {
202 // Set memory reference which is the best from all memory operations
203 // to be used for alignment. The pre-loop trip count is modified to align
204 // this reference to a vector-aligned address.
205 best_align_to_mem_ref = mem_ref;
206 best_iv_adjustment = iv_adjustment;
207 }
208
209 SWPointer align_to_ref_p(mem_ref, this);
210 // Set alignment relative to "align_to_ref" for all related memory operations.
211 for (int i = memops.size() - 1; i >= 0; i--) {
212 MemNode* s = memops.at(i)->as_Mem();
213 if (isomorphic(s, mem_ref)) {
214 SWPointer p2(s, this);
215 if (p2.comparable(align_to_ref_p)) {
216 int align = memory_alignment(s, iv_adjustment);
217 set_alignment(s, align);
218 }
219 }
220 }
221
222 // Create initial pack pairs of memory operations for which
223 // alignment is set and vectors will be aligned.
224 bool create_pack = true;
225 if (memory_alignment(mem_ref, best_iv_adjustment) == 0) {
226 if (!Matcher::misaligned_vectors_ok()) {
227 int vw = vector_width(mem_ref);
228 int vw_best = vector_width(best_align_to_mem_ref);
229 if (vw > vw_best) {
230 // Do not vectorize a memory access with more elements per vector
231 // if unaligned memory access is not allowed because number of
232 // iterations in pre-loop will be not enough to align it.
233 create_pack = false;
234 }
235 }
236 } else {
237 if (same_velt_type(mem_ref, best_align_to_mem_ref)) {
238 // Can't allow vectorization of unaligned memory accesses with the
239 // same type since it could be overlapped accesses to the same array.
240 create_pack = false;
241 } else {
242 // Allow independent (different type) unaligned memory operations
243 // if HW supports them.
244 if (!Matcher::misaligned_vectors_ok()) {
245 create_pack = false;
246 } else {
247 // Check if packs of the same memory type but
248 // with a different alignment were created before.
249 for (uint i = 0; i < align_to_refs.size(); i++) {
250 MemNode* mr = align_to_refs.at(i)->as_Mem();
251 if (same_velt_type(mr, mem_ref) &&
252 memory_alignment(mr, iv_adjustment) != 0)
253 create_pack = false;
254 }
255 }
256 }
257 }
258 if (create_pack) {
259 for (uint i = 0; i < memops.size(); i++) {
260 Node* s1 = memops.at(i);
261 int align = alignment(s1);
262 if (align == top_align) continue;
263 for (uint j = 0; j < memops.size(); j++) {
264 Node* s2 = memops.at(j);
265 if (alignment(s2) == top_align) continue;
266 if (s1 != s2 && are_adjacent_refs(s1, s2)) {
267 if (stmts_can_pack(s1, s2, align)) {
268 Node_List* pair = new Node_List();
269 pair->push(s1);
270 pair->push(s2);
271 _packset.append(pair);
272 }
273 }
274 }
275 }
276 } else { // Don't create unaligned pack
277 // First, remove remaining memory ops of the same type from the list.
278 for (int i = memops.size() - 1; i >= 0; i--) {
279 MemNode* s = memops.at(i)->as_Mem();
280 if (same_velt_type(s, mem_ref)) {
281 memops.remove(i);
282 }
283 }
284
285 // Second, remove already constructed packs of the same type.
286 for (int i = _packset.length() - 1; i >= 0; i--) {
287 Node_List* p = _packset.at(i);
288 MemNode* s = p->at(0)->as_Mem();
289 if (same_velt_type(s, mem_ref)) {
290 remove_pack_at(i);
291 }
292 }
293
294 // If needed find the best memory reference for loop alignment again.
295 if (same_velt_type(mem_ref, best_align_to_mem_ref)) {
296 // Put memory ops from remaining packs back on memops list for
297 // the best alignment search.
298 uint orig_msize = memops.size();
299 for (int i = 0; i < _packset.length(); i++) {
300 Node_List* p = _packset.at(i);
301 MemNode* s = p->at(0)->as_Mem();
302 assert(!same_velt_type(s, mem_ref), "sanity");
303 memops.push(s);
304 }
305 MemNode* best_align_to_mem_ref = find_align_to_ref(memops);
306 if (best_align_to_mem_ref == NULL) break;
307 best_iv_adjustment = get_iv_adjustment(best_align_to_mem_ref);
308 // Restore list.
309 while (memops.size() > orig_msize)
310 (void)memops.pop();
311 }
312 } // unaligned memory accesses
313
314 // Remove used mem nodes.
315 for (int i = memops.size() - 1; i >= 0; i--) {
316 MemNode* m = memops.at(i)->as_Mem();
317 if (alignment(m) != top_align) {
318 memops.remove(i);
319 }
320 }
321
322 } // while (memops.size() != 0
323 set_align_to_ref(best_align_to_mem_ref);
324
325 #ifndef PRODUCT
326 if (TraceSuperWord) {
327 tty->print_cr("\nAfter find_adjacent_refs");
328 print_packset();
329 }
330 #endif
331 }
332
333 //------------------------------find_align_to_ref---------------------------
334 // Find a memory reference to align the loop induction variable to.
335 // Looks first at stores then at loads, looking for a memory reference
336 // with the largest number of references similar to it.
337 MemNode* SuperWord::find_align_to_ref(Node_List &memops) {
338 GrowableArray<int> cmp_ct(arena(), memops.size(), memops.size(), 0);
339
340 // Count number of comparable memory ops
341 for (uint i = 0; i < memops.size(); i++) {
342 MemNode* s1 = memops.at(i)->as_Mem();
343 SWPointer p1(s1, this);
344 // Discard if pre loop can't align this reference
345 if (!ref_is_alignable(p1)) {
346 *cmp_ct.adr_at(i) = 0;
347 continue;
348 }
349 for (uint j = i+1; j < memops.size(); j++) {
350 MemNode* s2 = memops.at(j)->as_Mem();
351 if (isomorphic(s1, s2)) {
352 SWPointer p2(s2, this);
353 if (p1.comparable(p2)) {
354 (*cmp_ct.adr_at(i))++;
355 (*cmp_ct.adr_at(j))++;
356 }
357 }
358 }
359 }
360
361 // Find Store (or Load) with the greatest number of "comparable" references,
362 // biggest vector size, smallest data size and smallest iv offset.
363 int max_ct = 0;
364 int max_vw = 0;
365 int max_idx = -1;
366 int min_size = max_jint;
367 int min_iv_offset = max_jint;
368 for (uint j = 0; j < memops.size(); j++) {
369 MemNode* s = memops.at(j)->as_Mem();
370 if (s->is_Store()) {
371 int vw = vector_width_in_bytes(s);
372 assert(vw > 1, "sanity");
373 SWPointer p(s, this);
374 if (cmp_ct.at(j) > max_ct ||
375 cmp_ct.at(j) == max_ct &&
376 (vw > max_vw ||
377 vw == max_vw &&
378 (data_size(s) < min_size ||
379 data_size(s) == min_size &&
380 (p.offset_in_bytes() < min_iv_offset)))) {
381 max_ct = cmp_ct.at(j);
382 max_vw = vw;
383 max_idx = j;
384 min_size = data_size(s);
385 min_iv_offset = p.offset_in_bytes();
386 }
387 }
388 }
389 // If no stores, look at loads
390 if (max_ct == 0) {
391 for (uint j = 0; j < memops.size(); j++) {
392 MemNode* s = memops.at(j)->as_Mem();
393 if (s->is_Load()) {
394 int vw = vector_width_in_bytes(s);
395 assert(vw > 1, "sanity");
396 SWPointer p(s, this);
397 if (cmp_ct.at(j) > max_ct ||
398 cmp_ct.at(j) == max_ct &&
399 (vw > max_vw ||
400 vw == max_vw &&
401 (data_size(s) < min_size ||
402 data_size(s) == min_size &&
403 (p.offset_in_bytes() < min_iv_offset)))) {
404 max_ct = cmp_ct.at(j);
405 max_vw = vw;
406 max_idx = j;
407 min_size = data_size(s);
408 min_iv_offset = p.offset_in_bytes();
409 }
410 }
411 }
412 }
413
414 #ifdef ASSERT
415 if (TraceSuperWord && Verbose) {
416 tty->print_cr("\nVector memops after find_align_to_refs");
417 for (uint i = 0; i < memops.size(); i++) {
418 MemNode* s = memops.at(i)->as_Mem();
419 s->dump();
420 }
421 }
422 #endif
423
424 if (max_ct > 0) {
425 #ifdef ASSERT
426 if (TraceSuperWord) {
427 tty->print("\nVector align to node: ");
428 memops.at(max_idx)->as_Mem()->dump();
429 }
430 #endif
431 return memops.at(max_idx)->as_Mem();
432 }
433 return NULL;
434 }
435
436 //------------------------------ref_is_alignable---------------------------
437 // Can the preloop align the reference to position zero in the vector?
438 bool SuperWord::ref_is_alignable(SWPointer& p) {
439 if (!p.has_iv()) {
440 return true; // no induction variable
441 }
442 CountedLoopEndNode* pre_end = get_pre_loop_end(lp()->as_CountedLoop());
443 assert(pre_end->stride_is_con(), "pre loop stride is constant");
444 int preloop_stride = pre_end->stride_con();
445
446 int span = preloop_stride * p.scale_in_bytes();
447
448 // Stride one accesses are alignable.
449 if (ABS(span) == p.memory_size())
450 return true;
451
452 // If initial offset from start of object is computable,
453 // compute alignment within the vector.
454 int vw = vector_width_in_bytes(p.mem());
455 assert(vw > 1, "sanity");
456 if (vw % span == 0) {
457 Node* init_nd = pre_end->init_trip();
458 if (init_nd->is_Con() && p.invar() == NULL) {
459 int init = init_nd->bottom_type()->is_int()->get_con();
460
461 int init_offset = init * p.scale_in_bytes() + p.offset_in_bytes();
462 assert(init_offset >= 0, "positive offset from object start");
463
464 if (span > 0) {
465 return (vw - (init_offset % vw)) % span == 0;
466 } else {
467 assert(span < 0, "nonzero stride * scale");
468 return (init_offset % vw) % -span == 0;
469 }
470 }
471 }
472 return false;
473 }
474
475 //---------------------------get_iv_adjustment---------------------------
476 // Calculate loop's iv adjustment for this memory ops.
477 int SuperWord::get_iv_adjustment(MemNode* mem_ref) {
478 SWPointer align_to_ref_p(mem_ref, this);
479 int offset = align_to_ref_p.offset_in_bytes();
480 int scale = align_to_ref_p.scale_in_bytes();
481 int vw = vector_width_in_bytes(mem_ref);
482 assert(vw > 1, "sanity");
483 int stride_sign = (scale * iv_stride()) > 0 ? 1 : -1;
484 int iv_adjustment = (stride_sign * vw - (offset % vw)) % vw;
485
486 #ifndef PRODUCT
487 if (TraceSuperWord)
488 tty->print_cr("\noffset = %d iv_adjust = %d elt_size = %d scale = %d iv_stride = %d vect_size %d",
489 offset, iv_adjustment, align_to_ref_p.memory_size(), scale, iv_stride(), vw);
490 #endif
491 return iv_adjustment;
492 }
493
494 //---------------------------dependence_graph---------------------------
495 // Construct dependency graph.
496 // Add dependence edges to load/store nodes for memory dependence
497 // A.out()->DependNode.in(1) and DependNode.out()->B.prec(x)
498 void SuperWord::dependence_graph() {
499 // First, assign a dependence node to each memory node
500 for (int i = 0; i < _block.length(); i++ ) {
501 Node *n = _block.at(i);
502 if (n->is_Mem() || n->is_Phi() && n->bottom_type() == Type::MEMORY) {
503 _dg.make_node(n);
504 }
505 }
506
507 // For each memory slice, create the dependences
508 for (int i = 0; i < _mem_slice_head.length(); i++) {
509 Node* n = _mem_slice_head.at(i);
510 Node* n_tail = _mem_slice_tail.at(i);
511
512 // Get slice in predecessor order (last is first)
513 mem_slice_preds(n_tail, n, _nlist);
514
515 // Make the slice dependent on the root
516 DepMem* slice = _dg.dep(n);
517 _dg.make_edge(_dg.root(), slice);
518
519 // Create a sink for the slice
520 DepMem* slice_sink = _dg.make_node(NULL);
521 _dg.make_edge(slice_sink, _dg.tail());
522
523 // Now visit each pair of memory ops, creating the edges
524 for (int j = _nlist.length() - 1; j >= 0 ; j--) {
525 Node* s1 = _nlist.at(j);
526
527 // If no dependency yet, use slice
528 if (_dg.dep(s1)->in_cnt() == 0) {
529 _dg.make_edge(slice, s1);
530 }
531 SWPointer p1(s1->as_Mem(), this);
532 bool sink_dependent = true;
533 for (int k = j - 1; k >= 0; k--) {
534 Node* s2 = _nlist.at(k);
535 if (s1->is_Load() && s2->is_Load())
536 continue;
537 SWPointer p2(s2->as_Mem(), this);
538
539 int cmp = p1.cmp(p2);
540 if (SuperWordRTDepCheck &&
541 p1.base() != p2.base() && p1.valid() && p2.valid()) {
542 // Create a runtime check to disambiguate
543 OrderedPair pp(p1.base(), p2.base());
544 _disjoint_ptrs.append_if_missing(pp);
545 } else if (!SWPointer::not_equal(cmp)) {
546 // Possibly same address
547 _dg.make_edge(s1, s2);
548 sink_dependent = false;
549 }
550 }
551 if (sink_dependent) {
552 _dg.make_edge(s1, slice_sink);
553 }
554 }
555 #ifndef PRODUCT
556 if (TraceSuperWord) {
557 tty->print_cr("\nDependence graph for slice: %d", n->_idx);
558 for (int q = 0; q < _nlist.length(); q++) {
559 _dg.print(_nlist.at(q));
560 }
561 tty->cr();
562 }
563 #endif
564 _nlist.clear();
565 }
566
567 #ifndef PRODUCT
568 if (TraceSuperWord) {
569 tty->print_cr("\ndisjoint_ptrs: %s", _disjoint_ptrs.length() > 0 ? "" : "NONE");
570 for (int r = 0; r < _disjoint_ptrs.length(); r++) {
571 _disjoint_ptrs.at(r).print();
572 tty->cr();
573 }
574 tty->cr();
575 }
576 #endif
577 }
578
579 //---------------------------mem_slice_preds---------------------------
580 // Return a memory slice (node list) in predecessor order starting at "start"
581 void SuperWord::mem_slice_preds(Node* start, Node* stop, GrowableArray<Node*> &preds) {
582 assert(preds.length() == 0, "start empty");
583 Node* n = start;
584 Node* prev = NULL;
585 while (true) {
586 assert(in_bb(n), "must be in block");
587 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
588 Node* out = n->fast_out(i);
589 if (out->is_Load()) {
590 if (in_bb(out)) {
591 preds.push(out);
592 }
593 } else {
594 // FIXME
595 if (out->is_MergeMem() && !in_bb(out)) {
596 // Either unrolling is causing a memory edge not to disappear,
597 // or need to run igvn.optimize() again before SLP
598 } else if (out->is_Phi() && out->bottom_type() == Type::MEMORY && !in_bb(out)) {
599 // Ditto. Not sure what else to check further.
600 } else if (out->Opcode() == Op_StoreCM && out->in(MemNode::OopStore) == n) {
601 // StoreCM has an input edge used as a precedence edge.
602 // Maybe an issue when oop stores are vectorized.
603 } else {
604 assert(out == prev || prev == NULL, "no branches off of store slice");
605 }
606 }
607 }
608 if (n == stop) break;
609 preds.push(n);
610 prev = n;
611 n = n->in(MemNode::Memory);
612 }
613 }
614
615 //------------------------------stmts_can_pack---------------------------
616 // Can s1 and s2 be in a pack with s1 immediately preceding s2 and
617 // s1 aligned at "align"
618 bool SuperWord::stmts_can_pack(Node* s1, Node* s2, int align) {
619
620 // Do not use superword for non-primitives
621 BasicType bt1 = velt_basic_type(s1);
622 BasicType bt2 = velt_basic_type(s2);
623 if(!is_java_primitive(bt1) || !is_java_primitive(bt2))
624 return false;
625 if (Matcher::max_vector_size(bt1) < 2) {
626 return false; // No vectors for this type
627 }
628
629 if (isomorphic(s1, s2)) {
630 if (independent(s1, s2)) {
631 if (!exists_at(s1, 0) && !exists_at(s2, 1)) {
632 if (!s1->is_Mem() || are_adjacent_refs(s1, s2)) {
633 int s1_align = alignment(s1);
634 int s2_align = alignment(s2);
635 if (s1_align == top_align || s1_align == align) {
636 if (s2_align == top_align || s2_align == align + data_size(s1)) {
637 return true;
638 }
639 }
640 }
641 }
642 }
643 }
644 return false;
645 }
646
647 //------------------------------exists_at---------------------------
648 // Does s exist in a pack at position pos?
649 bool SuperWord::exists_at(Node* s, uint pos) {
650 for (int i = 0; i < _packset.length(); i++) {
651 Node_List* p = _packset.at(i);
652 if (p->at(pos) == s) {
653 return true;
654 }
655 }
656 return false;
657 }
658
659 //------------------------------are_adjacent_refs---------------------------
660 // Is s1 immediately before s2 in memory?
661 bool SuperWord::are_adjacent_refs(Node* s1, Node* s2) {
662 if (!s1->is_Mem() || !s2->is_Mem()) return false;
663 if (!in_bb(s1) || !in_bb(s2)) return false;
664
665 // Do not use superword for non-primitives
666 if (!is_java_primitive(s1->as_Mem()->memory_type()) ||
667 !is_java_primitive(s2->as_Mem()->memory_type())) {
668 return false;
669 }
670
671 // FIXME - co_locate_pack fails on Stores in different mem-slices, so
672 // only pack memops that are in the same alias set until that's fixed.
673 if (_phase->C->get_alias_index(s1->as_Mem()->adr_type()) !=
674 _phase->C->get_alias_index(s2->as_Mem()->adr_type()))
675 return false;
676 SWPointer p1(s1->as_Mem(), this);
677 SWPointer p2(s2->as_Mem(), this);
678 if (p1.base() != p2.base() || !p1.comparable(p2)) return false;
679 int diff = p2.offset_in_bytes() - p1.offset_in_bytes();
680 return diff == data_size(s1);
681 }
682
683 //------------------------------isomorphic---------------------------
684 // Are s1 and s2 similar?
685 bool SuperWord::isomorphic(Node* s1, Node* s2) {
686 if (s1->Opcode() != s2->Opcode()) return false;
687 if (s1->req() != s2->req()) return false;
688 if (s1->in(0) != s2->in(0)) return false;
689 if (!same_velt_type(s1, s2)) return false;
690 return true;
691 }
692
693 //------------------------------independent---------------------------
694 // Is there no data path from s1 to s2 or s2 to s1?
695 bool SuperWord::independent(Node* s1, Node* s2) {
696 // assert(s1->Opcode() == s2->Opcode(), "check isomorphic first");
697 int d1 = depth(s1);
698 int d2 = depth(s2);
699 if (d1 == d2) return s1 != s2;
700 Node* deep = d1 > d2 ? s1 : s2;
701 Node* shallow = d1 > d2 ? s2 : s1;
702
703 visited_clear();
704
705 return independent_path(shallow, deep);
706 }
707
708 //------------------------------independent_path------------------------------
709 // Helper for independent
710 bool SuperWord::independent_path(Node* shallow, Node* deep, uint dp) {
711 if (dp >= 1000) return false; // stop deep recursion
712 visited_set(deep);
713 int shal_depth = depth(shallow);
714 assert(shal_depth <= depth(deep), "must be");
715 for (DepPreds preds(deep, _dg); !preds.done(); preds.next()) {
716 Node* pred = preds.current();
717 if (in_bb(pred) && !visited_test(pred)) {
718 if (shallow == pred) {
719 return false;
720 }
721 if (shal_depth < depth(pred) && !independent_path(shallow, pred, dp+1)) {
722 return false;
723 }
724 }
725 }
726 return true;
727 }
728
729 //------------------------------set_alignment---------------------------
730 void SuperWord::set_alignment(Node* s1, Node* s2, int align) {
731 set_alignment(s1, align);
732 if (align == top_align || align == bottom_align) {
733 set_alignment(s2, align);
734 } else {
735 set_alignment(s2, align + data_size(s1));
736 }
737 }
738
739 //------------------------------data_size---------------------------
740 int SuperWord::data_size(Node* s) {
741 int bsize = type2aelembytes(velt_basic_type(s));
742 assert(bsize != 0, "valid size");
743 return bsize;
744 }
745
746 //------------------------------extend_packlist---------------------------
747 // Extend packset by following use->def and def->use links from pack members.
748 void SuperWord::extend_packlist() {
749 bool changed;
750 do {
751 changed = false;
752 for (int i = 0; i < _packset.length(); i++) {
753 Node_List* p = _packset.at(i);
754 changed |= follow_use_defs(p);
755 changed |= follow_def_uses(p);
756 }
757 } while (changed);
758
759 #ifndef PRODUCT
760 if (TraceSuperWord) {
761 tty->print_cr("\nAfter extend_packlist");
762 print_packset();
763 }
764 #endif
765 }
766
767 //------------------------------follow_use_defs---------------------------
768 // Extend the packset by visiting operand definitions of nodes in pack p
769 bool SuperWord::follow_use_defs(Node_List* p) {
770 assert(p->size() == 2, "just checking");
771 Node* s1 = p->at(0);
772 Node* s2 = p->at(1);
773 assert(s1->req() == s2->req(), "just checking");
774 assert(alignment(s1) + data_size(s1) == alignment(s2), "just checking");
775
776 if (s1->is_Load()) return false;
777
778 int align = alignment(s1);
779 bool changed = false;
780 int start = s1->is_Store() ? MemNode::ValueIn : 1;
781 int end = s1->is_Store() ? MemNode::ValueIn+1 : s1->req();
782 for (int j = start; j < end; j++) {
783 Node* t1 = s1->in(j);
784 Node* t2 = s2->in(j);
785 if (!in_bb(t1) || !in_bb(t2))
786 continue;
787 if (stmts_can_pack(t1, t2, align)) {
788 if (est_savings(t1, t2) >= 0) {
789 Node_List* pair = new Node_List();
790 pair->push(t1);
791 pair->push(t2);
792 _packset.append(pair);
793 set_alignment(t1, t2, align);
794 changed = true;
795 }
796 }
797 }
798 return changed;
799 }
800
801 //------------------------------follow_def_uses---------------------------
802 // Extend the packset by visiting uses of nodes in pack p
803 bool SuperWord::follow_def_uses(Node_List* p) {
804 bool changed = false;
805 Node* s1 = p->at(0);
806 Node* s2 = p->at(1);
807 assert(p->size() == 2, "just checking");
808 assert(s1->req() == s2->req(), "just checking");
809 assert(alignment(s1) + data_size(s1) == alignment(s2), "just checking");
810
811 if (s1->is_Store()) return false;
812
813 int align = alignment(s1);
814 int savings = -1;
815 Node* u1 = NULL;
816 Node* u2 = NULL;
817 for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
818 Node* t1 = s1->fast_out(i);
819 if (!in_bb(t1)) continue;
820 for (DUIterator_Fast jmax, j = s2->fast_outs(jmax); j < jmax; j++) {
821 Node* t2 = s2->fast_out(j);
822 if (!in_bb(t2)) continue;
823 if (!opnd_positions_match(s1, t1, s2, t2))
824 continue;
825 if (stmts_can_pack(t1, t2, align)) {
826 int my_savings = est_savings(t1, t2);
827 if (my_savings > savings) {
828 savings = my_savings;
829 u1 = t1;
830 u2 = t2;
831 }
832 }
833 }
834 }
835 if (savings >= 0) {
836 Node_List* pair = new Node_List();
837 pair->push(u1);
838 pair->push(u2);
839 _packset.append(pair);
840 set_alignment(u1, u2, align);
841 changed = true;
842 }
843 return changed;
844 }
845
846 //---------------------------opnd_positions_match-------------------------
847 // Is the use of d1 in u1 at the same operand position as d2 in u2?
848 bool SuperWord::opnd_positions_match(Node* d1, Node* u1, Node* d2, Node* u2) {
849 uint ct = u1->req();
850 if (ct != u2->req()) return false;
851 uint i1 = 0;
852 uint i2 = 0;
853 do {
854 for (i1++; i1 < ct; i1++) if (u1->in(i1) == d1) break;
855 for (i2++; i2 < ct; i2++) if (u2->in(i2) == d2) break;
856 if (i1 != i2) {
857 if ((i1 == (3-i2)) && (u2->is_Add() || u2->is_Mul())) {
858 // Further analysis relies on operands position matching.
859 u2->swap_edges(i1, i2);
860 } else {
861 return false;
862 }
863 }
864 } while (i1 < ct);
865 return true;
866 }
867
868 //------------------------------est_savings---------------------------
869 // Estimate the savings from executing s1 and s2 as a pack
870 int SuperWord::est_savings(Node* s1, Node* s2) {
871 int save_in = 2 - 1; // 2 operations per instruction in packed form
872
873 // inputs
874 for (uint i = 1; i < s1->req(); i++) {
875 Node* x1 = s1->in(i);
876 Node* x2 = s2->in(i);
877 if (x1 != x2) {
878 if (are_adjacent_refs(x1, x2)) {
879 save_in += adjacent_profit(x1, x2);
880 } else if (!in_packset(x1, x2)) {
881 save_in -= pack_cost(2);
882 } else {
883 save_in += unpack_cost(2);
884 }
885 }
886 }
887
888 // uses of result
889 uint ct = 0;
890 int save_use = 0;
891 for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
892 Node* s1_use = s1->fast_out(i);
893 for (int j = 0; j < _packset.length(); j++) {
894 Node_List* p = _packset.at(j);
895 if (p->at(0) == s1_use) {
896 for (DUIterator_Fast kmax, k = s2->fast_outs(kmax); k < kmax; k++) {
897 Node* s2_use = s2->fast_out(k);
898 if (p->at(p->size()-1) == s2_use) {
899 ct++;
900 if (are_adjacent_refs(s1_use, s2_use)) {
901 save_use += adjacent_profit(s1_use, s2_use);
902 }
903 }
904 }
905 }
906 }
907 }
908
909 if (ct < s1->outcnt()) save_use += unpack_cost(1);
910 if (ct < s2->outcnt()) save_use += unpack_cost(1);
911
912 return MAX2(save_in, save_use);
913 }
914
915 //------------------------------costs---------------------------
916 int SuperWord::adjacent_profit(Node* s1, Node* s2) { return 2; }
917 int SuperWord::pack_cost(int ct) { return ct; }
918 int SuperWord::unpack_cost(int ct) { return ct; }
919
920 //------------------------------combine_packs---------------------------
921 // Combine packs A and B with A.last == B.first into A.first..,A.last,B.second,..B.last
922 void SuperWord::combine_packs() {
923 bool changed = true;
924 // Combine packs regardless max vector size.
925 while (changed) {
926 changed = false;
927 for (int i = 0; i < _packset.length(); i++) {
928 Node_List* p1 = _packset.at(i);
929 if (p1 == NULL) continue;
930 for (int j = 0; j < _packset.length(); j++) {
931 Node_List* p2 = _packset.at(j);
932 if (p2 == NULL) continue;
933 if (i == j) continue;
934 if (p1->at(p1->size()-1) == p2->at(0)) {
935 for (uint k = 1; k < p2->size(); k++) {
936 p1->push(p2->at(k));
937 }
938 _packset.at_put(j, NULL);
939 changed = true;
940 }
941 }
942 }
943 }
944
945 // Split packs which have size greater then max vector size.
946 for (int i = 0; i < _packset.length(); i++) {
947 Node_List* p1 = _packset.at(i);
948 if (p1 != NULL) {
949 BasicType bt = velt_basic_type(p1->at(0));
950 uint max_vlen = Matcher::max_vector_size(bt); // Max elements in vector
951 assert(is_power_of_2(max_vlen), "sanity");
952 uint psize = p1->size();
953 if (!is_power_of_2(psize)) {
954 // Skip pack which can't be vector.
955 // case1: for(...) { a[i] = i; } elements values are different (i+x)
956 // case2: for(...) { a[i] = b[i+1]; } can't align both, load and store
957 _packset.at_put(i, NULL);
958 continue;
959 }
960 if (psize > max_vlen) {
961 Node_List* pack = new Node_List();
962 for (uint j = 0; j < psize; j++) {
963 pack->push(p1->at(j));
964 if (pack->size() >= max_vlen) {
965 assert(is_power_of_2(pack->size()), "sanity");
966 _packset.append(pack);
967 pack = new Node_List();
968 }
969 }
970 _packset.at_put(i, NULL);
971 }
972 }
973 }
974
975 // Compress list.
976 for (int i = _packset.length() - 1; i >= 0; i--) {
977 Node_List* p1 = _packset.at(i);
978 if (p1 == NULL) {
979 _packset.remove_at(i);
980 }
981 }
982
983 #ifndef PRODUCT
984 if (TraceSuperWord) {
985 tty->print_cr("\nAfter combine_packs");
986 print_packset();
987 }
988 #endif
989 }
990
991 //-----------------------------construct_my_pack_map--------------------------
992 // Construct the map from nodes to packs. Only valid after the
993 // point where a node is only in one pack (after combine_packs).
994 void SuperWord::construct_my_pack_map() {
995 Node_List* rslt = NULL;
996 for (int i = 0; i < _packset.length(); i++) {
997 Node_List* p = _packset.at(i);
998 for (uint j = 0; j < p->size(); j++) {
999 Node* s = p->at(j);
1000 assert(my_pack(s) == NULL, "only in one pack");
1001 set_my_pack(s, p);
1002 }
1003 }
1004 }
1005
1006 //------------------------------filter_packs---------------------------
1007 // Remove packs that are not implemented or not profitable.
1008 void SuperWord::filter_packs() {
1009
1010 // Remove packs that are not implemented
1011 for (int i = _packset.length() - 1; i >= 0; i--) {
1012 Node_List* pk = _packset.at(i);
1013 bool impl = implemented(pk);
1014 if (!impl) {
1015 #ifndef PRODUCT
1016 if (TraceSuperWord && Verbose) {
1017 tty->print_cr("Unimplemented");
1018 pk->at(0)->dump();
1019 }
1020 #endif
1021 remove_pack_at(i);
1022 }
1023 }
1024
1025 // Remove packs that are not profitable
1026 bool changed;
1027 do {
1028 changed = false;
1029 for (int i = _packset.length() - 1; i >= 0; i--) {
1030 Node_List* pk = _packset.at(i);
1031 bool prof = profitable(pk);
1032 if (!prof) {
1033 #ifndef PRODUCT
1034 if (TraceSuperWord && Verbose) {
1035 tty->print_cr("Unprofitable");
1036 pk->at(0)->dump();
1037 }
1038 #endif
1039 remove_pack_at(i);
1040 changed = true;
1041 }
1042 }
1043 } while (changed);
1044
1045 #ifndef PRODUCT
1046 if (TraceSuperWord) {
1047 tty->print_cr("\nAfter filter_packs");
1048 print_packset();
1049 tty->cr();
1050 }
1051 #endif
1052 }
1053
1054 //------------------------------implemented---------------------------
1055 // Can code be generated for pack p?
1056 bool SuperWord::implemented(Node_List* p) {
1057 Node* p0 = p->at(0);
1058 return VectorNode::implemented(p0->Opcode(), p->size(), velt_basic_type(p0));
1059 }
1060
1061 //------------------------------profitable---------------------------
1062 // For pack p, are all operands and all uses (with in the block) vector?
1063 bool SuperWord::profitable(Node_List* p) {
1064 Node* p0 = p->at(0);
1065 uint start, end;
1066 vector_opd_range(p0, &start, &end);
1067
1068 // Return false if some input is not vector and inside block
1069 for (uint i = start; i < end; i++) {
1070 if (!is_vector_use(p0, i)) {
1071 // For now, return false if not scalar promotion case (inputs are the same.)
1072 // Later, implement PackNode and allow differing, non-vector inputs
1073 // (maybe just the ones from outside the block.)
1074 Node* p0_def = p0->in(i);
1075 for (uint j = 1; j < p->size(); j++) {
1076 Node* use = p->at(j);
1077 Node* def = use->in(i);
1078 if (p0_def != def)
1079 return false;
1080 }
1081 }
1082 }
1083 if (!p0->is_Store()) {
1084 // For now, return false if not all uses are vector.
1085 // Later, implement ExtractNode and allow non-vector uses (maybe
1086 // just the ones outside the block.)
1087 for (uint i = 0; i < p->size(); i++) {
1088 Node* def = p->at(i);
1089 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
1090 Node* use = def->fast_out(j);
1091 for (uint k = 0; k < use->req(); k++) {
1092 Node* n = use->in(k);
1093 if (def == n) {
1094 if (!is_vector_use(use, k)) {
1095 return false;
1096 }
1097 }
1098 }
1099 }
1100 }
1101 }
1102 return true;
1103 }
1104
1105 //------------------------------schedule---------------------------
1106 // Adjust the memory graph for the packed operations
1107 void SuperWord::schedule() {
1108
1109 // Co-locate in the memory graph the members of each memory pack
1110 for (int i = 0; i < _packset.length(); i++) {
1111 co_locate_pack(_packset.at(i));
1112 }
1113 }
1114
1115 //-------------------------------remove_and_insert-------------------
1116 // Remove "current" from its current position in the memory graph and insert
1117 // it after the appropriate insertion point (lip or uip).
1118 void SuperWord::remove_and_insert(MemNode *current, MemNode *prev, MemNode *lip,
1119 Node *uip, Unique_Node_List &sched_before) {
1120 Node* my_mem = current->in(MemNode::Memory);
1121 bool sched_up = sched_before.member(current);
1122
1123 // remove current_store from its current position in the memmory graph
1124 for (DUIterator i = current->outs(); current->has_out(i); i++) {
1125 Node* use = current->out(i);
1126 if (use->is_Mem()) {
1127 assert(use->in(MemNode::Memory) == current, "must be");
1128 if (use == prev) { // connect prev to my_mem
1129 _igvn.replace_input_of(use, MemNode::Memory, my_mem);
1130 --i; //deleted this edge; rescan position
1131 } else if (sched_before.member(use)) {
1132 if (!sched_up) { // Will be moved together with current
1133 _igvn.replace_input_of(use, MemNode::Memory, uip);
1134 --i; //deleted this edge; rescan position
1135 }
1136 } else {
1137 if (sched_up) { // Will be moved together with current
1138 _igvn.replace_input_of(use, MemNode::Memory, lip);
1139 --i; //deleted this edge; rescan position
1140 }
1141 }
1142 }
1143 }
1144
1145 Node *insert_pt = sched_up ? uip : lip;
1146
1147 // all uses of insert_pt's memory state should use current's instead
1148 for (DUIterator i = insert_pt->outs(); insert_pt->has_out(i); i++) {
1149 Node* use = insert_pt->out(i);
1150 if (use->is_Mem()) {
1151 assert(use->in(MemNode::Memory) == insert_pt, "must be");
1152 _igvn.replace_input_of(use, MemNode::Memory, current);
1153 --i; //deleted this edge; rescan position
1154 } else if (!sched_up && use->is_Phi() && use->bottom_type() == Type::MEMORY) {
1155 uint pos; //lip (lower insert point) must be the last one in the memory slice
1156 for (pos=1; pos < use->req(); pos++) {
1157 if (use->in(pos) == insert_pt) break;
1158 }
1159 _igvn.replace_input_of(use, pos, current);
1160 --i;
1161 }
1162 }
1163
1164 //connect current to insert_pt
1165 _igvn.replace_input_of(current, MemNode::Memory, insert_pt);
1166 }
1167
1168 //------------------------------co_locate_pack----------------------------------
1169 // To schedule a store pack, we need to move any sandwiched memory ops either before
1170 // or after the pack, based upon dependence information:
1171 // (1) If any store in the pack depends on the sandwiched memory op, the
1172 // sandwiched memory op must be scheduled BEFORE the pack;
1173 // (2) If a sandwiched memory op depends on any store in the pack, the
1174 // sandwiched memory op must be scheduled AFTER the pack;
1175 // (3) If a sandwiched memory op (say, memA) depends on another sandwiched
1176 // memory op (say memB), memB must be scheduled before memA. So, if memA is
1177 // scheduled before the pack, memB must also be scheduled before the pack;
1178 // (4) If there is no dependence restriction for a sandwiched memory op, we simply
1179 // schedule this store AFTER the pack
1180 // (5) We know there is no dependence cycle, so there in no other case;
1181 // (6) Finally, all memory ops in another single pack should be moved in the same direction.
1182 //
1183 // To schedule a load pack, we use the memory state of either the first or the last load in
1184 // the pack, based on the dependence constraint.
1185 void SuperWord::co_locate_pack(Node_List* pk) {
1186 if (pk->at(0)->is_Store()) {
1187 MemNode* first = executed_first(pk)->as_Mem();
1188 MemNode* last = executed_last(pk)->as_Mem();
1189 Unique_Node_List schedule_before_pack;
1190 Unique_Node_List memops;
1191
1192 MemNode* current = last->in(MemNode::Memory)->as_Mem();
1193 MemNode* previous = last;
1194 while (true) {
1195 assert(in_bb(current), "stay in block");
1196 memops.push(previous);
1197 for (DUIterator i = current->outs(); current->has_out(i); i++) {
1198 Node* use = current->out(i);
1199 if (use->is_Mem() && use != previous)
1200 memops.push(use);
1201 }
1202 if (current == first) break;
1203 previous = current;
1204 current = current->in(MemNode::Memory)->as_Mem();
1205 }
1206
1207 // determine which memory operations should be scheduled before the pack
1208 for (uint i = 1; i < memops.size(); i++) {
1209 Node *s1 = memops.at(i);
1210 if (!in_pack(s1, pk) && !schedule_before_pack.member(s1)) {
1211 for (uint j = 0; j< i; j++) {
1212 Node *s2 = memops.at(j);
1213 if (!independent(s1, s2)) {
1214 if (in_pack(s2, pk) || schedule_before_pack.member(s2)) {
1215 schedule_before_pack.push(s1); // s1 must be scheduled before
1216 Node_List* mem_pk = my_pack(s1);
1217 if (mem_pk != NULL) {
1218 for (uint ii = 0; ii < mem_pk->size(); ii++) {
1219 Node* s = mem_pk->at(ii); // follow partner
1220 if (memops.member(s) && !schedule_before_pack.member(s))
1221 schedule_before_pack.push(s);
1222 }
1223 }
1224 break;
1225 }
1226 }
1227 }
1228 }
1229 }
1230
1231 Node* upper_insert_pt = first->in(MemNode::Memory);
1232 // Following code moves loads connected to upper_insert_pt below aliased stores.
1233 // Collect such loads here and reconnect them back to upper_insert_pt later.
1234 memops.clear();
1235 for (DUIterator i = upper_insert_pt->outs(); upper_insert_pt->has_out(i); i++) {
1236 Node* use = upper_insert_pt->out(i);
1237 if (!use->is_Store())
1238 memops.push(use);
1239 }
1240
1241 MemNode* lower_insert_pt = last;
1242 previous = last; //previous store in pk
1243 current = last->in(MemNode::Memory)->as_Mem();
1244
1245 // start scheduling from "last" to "first"
1246 while (true) {
1247 assert(in_bb(current), "stay in block");
1248 assert(in_pack(previous, pk), "previous stays in pack");
1249 Node* my_mem = current->in(MemNode::Memory);
1250
1251 if (in_pack(current, pk)) {
1252 // Forward users of my memory state (except "previous) to my input memory state
1253 for (DUIterator i = current->outs(); current->has_out(i); i++) {
1254 Node* use = current->out(i);
1255 if (use->is_Mem() && use != previous) {
1256 assert(use->in(MemNode::Memory) == current, "must be");
1257 if (schedule_before_pack.member(use)) {
1258 _igvn.replace_input_of(use, MemNode::Memory, upper_insert_pt);
1259 } else {
1260 _igvn.replace_input_of(use, MemNode::Memory, lower_insert_pt);
1261 }
1262 --i; // deleted this edge; rescan position
1263 }
1264 }
1265 previous = current;
1266 } else { // !in_pack(current, pk) ==> a sandwiched store
1267 remove_and_insert(current, previous, lower_insert_pt, upper_insert_pt, schedule_before_pack);
1268 }
1269
1270 if (current == first) break;
1271 current = my_mem->as_Mem();
1272 } // end while
1273
1274 // Reconnect loads back to upper_insert_pt.
1275 for (uint i = 0; i < memops.size(); i++) {
1276 Node *ld = memops.at(i);
1277 if (ld->in(MemNode::Memory) != upper_insert_pt) {
1278 _igvn.replace_input_of(ld, MemNode::Memory, upper_insert_pt);
1279 }
1280 }
1281 } else if (pk->at(0)->is_Load()) { //load
1282 // all loads in the pack should have the same memory state. By default,
1283 // we use the memory state of the last load. However, if any load could
1284 // not be moved down due to the dependence constraint, we use the memory
1285 // state of the first load.
1286 Node* last_mem = executed_last(pk)->in(MemNode::Memory);
1287 Node* first_mem = executed_first(pk)->in(MemNode::Memory);
1288 bool schedule_last = true;
1289 for (uint i = 0; i < pk->size(); i++) {
1290 Node* ld = pk->at(i);
1291 for (Node* current = last_mem; current != ld->in(MemNode::Memory);
1292 current=current->in(MemNode::Memory)) {
1293 assert(current != first_mem, "corrupted memory graph");
1294 if(current->is_Mem() && !independent(current, ld)){
1295 schedule_last = false; // a later store depends on this load
1296 break;
1297 }
1298 }
1299 }
1300
1301 Node* mem_input = schedule_last ? last_mem : first_mem;
1302 _igvn.hash_delete(mem_input);
1303 // Give each load the same memory state
1304 for (uint i = 0; i < pk->size(); i++) {
1305 LoadNode* ld = pk->at(i)->as_Load();
1306 _igvn.replace_input_of(ld, MemNode::Memory, mem_input);
1307 }
1308 }
1309 }
1310
1311 //------------------------------output---------------------------
1312 // Convert packs into vector node operations
1313 void SuperWord::output() {
1314 if (_packset.length() == 0) return;
1315
1316 #ifndef PRODUCT
1317 if (TraceLoopOpts) {
1318 tty->print("SuperWord ");
1319 lpt()->dump_head();
1320 }
1321 #endif
1322
1323 // MUST ENSURE main loop's initial value is properly aligned:
1324 // (iv_initial_value + min_iv_offset) % vector_width_in_bytes() == 0
1325
1326 align_initial_loop_index(align_to_ref());
1327
1328 // Insert extract (unpack) operations for scalar uses
1329 for (int i = 0; i < _packset.length(); i++) {
1330 insert_extracts(_packset.at(i));
1331 }
1332
1333 for (int i = 0; i < _block.length(); i++) {
1334 Node* n = _block.at(i);
1335 Node_List* p = my_pack(n);
1336 if (p && n == executed_last(p)) {
1337 uint vlen = p->size();
1338 Node* vn = NULL;
1339 Node* low_adr = p->at(0);
1340 Node* first = executed_first(p);
1341 int opc = n->Opcode();
1342 if (n->is_Load()) {
1343 Node* ctl = n->in(MemNode::Control);
1344 Node* mem = first->in(MemNode::Memory);
1345 Node* adr = low_adr->in(MemNode::Address);
1346 const TypePtr* atyp = n->adr_type();
1347 vn = LoadVectorNode::make(_phase->C, opc, ctl, mem, adr, atyp, vlen, velt_basic_type(n));
1348 } else if (n->is_Store()) {
1349 // Promote value to be stored to vector
1350 Node* val = vector_opd(p, MemNode::ValueIn);
1351 Node* ctl = n->in(MemNode::Control);
1352 Node* mem = first->in(MemNode::Memory);
1353 Node* adr = low_adr->in(MemNode::Address);
1354 const TypePtr* atyp = n->adr_type();
1355 vn = StoreVectorNode::make(_phase->C, opc, ctl, mem, adr, atyp, val, vlen);
1356 } else if (n->req() == 3) {
1357 // Promote operands to vector
1358 Node* in1 = vector_opd(p, 1);
1359 Node* in2 = vector_opd(p, 2);
1360 if (VectorNode::is_invariant_vector(in1) && (n->is_Add() || n->is_Mul())) {
1361 // Move invariant vector input into second position to avoid register spilling.
1362 Node* tmp = in1;
1363 in1 = in2;
1364 in2 = tmp;
1365 }
1366 vn = VectorNode::make(_phase->C, opc, in1, in2, vlen, velt_basic_type(n));
1367 } else {
1368 ShouldNotReachHere();
1369 }
1370 assert(vn != NULL, "sanity");
1371 _phase->_igvn.register_new_node_with_optimizer(vn);
1372 _phase->set_ctrl(vn, _phase->get_ctrl(p->at(0)));
1373 for (uint j = 0; j < p->size(); j++) {
1374 Node* pm = p->at(j);
1375 _igvn.replace_node(pm, vn);
1376 }
1377 _igvn._worklist.push(vn);
1378 #ifdef ASSERT
1379 if (TraceNewVectors) {
1380 tty->print("new Vector node: ");
1381 vn->dump();
1382 }
1383 #endif
1384 }
1385 }
1386 }
1387
1388 //------------------------------vector_opd---------------------------
1389 // Create a vector operand for the nodes in pack p for operand: in(opd_idx)
1390 Node* SuperWord::vector_opd(Node_List* p, int opd_idx) {
1391 Node* p0 = p->at(0);
1392 uint vlen = p->size();
1393 Node* opd = p0->in(opd_idx);
1394
1395 bool same_opd = true;
1396 for (uint i = 1; i < vlen; i++) {
1397 Node* pi = p->at(i);
1398 Node* in = pi->in(opd_idx);
1399 if (opd != in) {
1400 same_opd = false;
1401 break;
1402 }
1403 }
1404
1405 if (same_opd) {
1406 if (opd->is_Vector() || opd->is_LoadVector()) {
1407 return opd; // input is matching vector
1408 }
1409 if ((opd_idx == 2) && VectorNode::is_shift(p0)) {
1410 // No vector is needed for shift count.
1411 // Vector instructions do not mask shift count, do it here.
1412 Compile* C = _phase->C;
1413 Node* cnt = opd;
1414 juint mask = (p0->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
1415 const TypeInt* t = opd->find_int_type();
1416 if (t != NULL && t->is_con()) {
1417 juint shift = t->get_con();
1418 if (shift > mask) { // Unsigned cmp
1419 cnt = ConNode::make(C, TypeInt::make(shift & mask));
1420 }
1421 } else {
1422 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
1423 cnt = ConNode::make(C, TypeInt::make(mask));
1424 _phase->_igvn.register_new_node_with_optimizer(cnt);
1425 cnt = new (C, 3) AndINode(opd, cnt);
1426 _phase->_igvn.register_new_node_with_optimizer(cnt);
1427 _phase->set_ctrl(cnt, _phase->get_ctrl(opd));
1428 }
1429 assert(opd->bottom_type()->isa_int(), "int type only");
1430 // Move non constant shift count into XMM register.
1431 cnt = new (_phase->C, 2) MoveI2FNode(cnt);
1432 }
1433 if (cnt != opd) {
1434 _phase->_igvn.register_new_node_with_optimizer(cnt);
1435 _phase->set_ctrl(cnt, _phase->get_ctrl(opd));
1436 }
1437 return cnt;
1438 }
1439 assert(!opd->is_StoreVector(), "such vector is not expected here");
1440 // Convert scalar input to vector with the same number of elements as
1441 // p0's vector. Use p0's type because size of operand's container in
1442 // vector should match p0's size regardless operand's size.
1443 const Type* p0_t = velt_type(p0);
1444 VectorNode* vn = VectorNode::scalar2vector(_phase->C, opd, vlen, p0_t);
1445
1446 _phase->_igvn.register_new_node_with_optimizer(vn);
1447 _phase->set_ctrl(vn, _phase->get_ctrl(opd));
1448 #ifdef ASSERT
1449 if (TraceNewVectors) {
1450 tty->print("new Vector node: ");
1451 vn->dump();
1452 }
1453 #endif
1454 return vn;
1455 }
1456
1457 // Insert pack operation
1458 BasicType bt = velt_basic_type(p0);
1459 PackNode* pk = PackNode::make(_phase->C, opd, vlen, bt);
1460 DEBUG_ONLY( const BasicType opd_bt = opd->bottom_type()->basic_type(); )
1461
1462 for (uint i = 1; i < vlen; i++) {
1463 Node* pi = p->at(i);
1464 Node* in = pi->in(opd_idx);
1465 assert(my_pack(in) == NULL, "Should already have been unpacked");
1466 assert(opd_bt == in->bottom_type()->basic_type(), "all same type");
1467 pk->add_opd(i, in);
1468 }
1469 _phase->_igvn.register_new_node_with_optimizer(pk);
1470 _phase->set_ctrl(pk, _phase->get_ctrl(opd));
1471 #ifdef ASSERT
1472 if (TraceNewVectors) {
1473 tty->print("new Vector node: ");
1474 pk->dump();
1475 }
1476 #endif
1477 return pk;
1478 }
1479
1480 //------------------------------insert_extracts---------------------------
1481 // If a use of pack p is not a vector use, then replace the
1482 // use with an extract operation.
1483 void SuperWord::insert_extracts(Node_List* p) {
1484 if (p->at(0)->is_Store()) return;
1485 assert(_n_idx_list.is_empty(), "empty (node,index) list");
1486
1487 // Inspect each use of each pack member. For each use that is
1488 // not a vector use, replace the use with an extract operation.
1489
1490 for (uint i = 0; i < p->size(); i++) {
1491 Node* def = p->at(i);
1492 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
1493 Node* use = def->fast_out(j);
1494 for (uint k = 0; k < use->req(); k++) {
1495 Node* n = use->in(k);
1496 if (def == n) {
1497 if (!is_vector_use(use, k)) {
1498 _n_idx_list.push(use, k);
1499 }
1500 }
1501 }
1502 }
1503 }
1504
1505 while (_n_idx_list.is_nonempty()) {
1506 Node* use = _n_idx_list.node();
1507 int idx = _n_idx_list.index();
1508 _n_idx_list.pop();
1509 Node* def = use->in(idx);
1510
1511 // Insert extract operation
1512 _igvn.hash_delete(def);
1513 int def_pos = alignment(def) / data_size(def);
1514
1515 Node* ex = ExtractNode::make(_phase->C, def, def_pos, velt_basic_type(def));
1516 _phase->_igvn.register_new_node_with_optimizer(ex);
1517 _phase->set_ctrl(ex, _phase->get_ctrl(def));
1518 _igvn.replace_input_of(use, idx, ex);
1519 _igvn._worklist.push(def);
1520
1521 bb_insert_after(ex, bb_idx(def));
1522 set_velt_type(ex, velt_type(def));
1523 }
1524 }
1525
1526 //------------------------------is_vector_use---------------------------
1527 // Is use->in(u_idx) a vector use?
1528 bool SuperWord::is_vector_use(Node* use, int u_idx) {
1529 Node_List* u_pk = my_pack(use);
1530 if (u_pk == NULL) return false;
1531 Node* def = use->in(u_idx);
1532 Node_List* d_pk = my_pack(def);
1533 if (d_pk == NULL) {
1534 // check for scalar promotion
1535 Node* n = u_pk->at(0)->in(u_idx);
1536 for (uint i = 1; i < u_pk->size(); i++) {
1537 if (u_pk->at(i)->in(u_idx) != n) return false;
1538 }
1539 return true;
1540 }
1541 if (u_pk->size() != d_pk->size())
1542 return false;
1543 for (uint i = 0; i < u_pk->size(); i++) {
1544 Node* ui = u_pk->at(i);
1545 Node* di = d_pk->at(i);
1546 if (ui->in(u_idx) != di || alignment(ui) != alignment(di))
1547 return false;
1548 }
1549 return true;
1550 }
1551
1552 //------------------------------construct_bb---------------------------
1553 // Construct reverse postorder list of block members
1554 void SuperWord::construct_bb() {
1555 Node* entry = bb();
1556
1557 assert(_stk.length() == 0, "stk is empty");
1558 assert(_block.length() == 0, "block is empty");
1559 assert(_data_entry.length() == 0, "data_entry is empty");
1560 assert(_mem_slice_head.length() == 0, "mem_slice_head is empty");
1561 assert(_mem_slice_tail.length() == 0, "mem_slice_tail is empty");
1562
1563 // Find non-control nodes with no inputs from within block,
1564 // create a temporary map from node _idx to bb_idx for use
1565 // by the visited and post_visited sets,
1566 // and count number of nodes in block.
1567 int bb_ct = 0;
1568 for (uint i = 0; i < lpt()->_body.size(); i++ ) {
1569 Node *n = lpt()->_body.at(i);
1570 set_bb_idx(n, i); // Create a temporary map
1571 if (in_bb(n)) {
1572 bb_ct++;
1573 if (!n->is_CFG()) {
1574 bool found = false;
1575 for (uint j = 0; j < n->req(); j++) {
1576 Node* def = n->in(j);
1577 if (def && in_bb(def)) {
1578 found = true;
1579 break;
1580 }
1581 }
1582 if (!found) {
1583 assert(n != entry, "can't be entry");
1584 _data_entry.push(n);
1585 }
1586 }
1587 }
1588 }
1589
1590 // Find memory slices (head and tail)
1591 for (DUIterator_Fast imax, i = lp()->fast_outs(imax); i < imax; i++) {
1592 Node *n = lp()->fast_out(i);
1593 if (in_bb(n) && (n->is_Phi() && n->bottom_type() == Type::MEMORY)) {
1594 Node* n_tail = n->in(LoopNode::LoopBackControl);
1595 if (n_tail != n->in(LoopNode::EntryControl)) {
1596 _mem_slice_head.push(n);
1597 _mem_slice_tail.push(n_tail);
1598 }
1599 }
1600 }
1601
1602 // Create an RPO list of nodes in block
1603
1604 visited_clear();
1605 post_visited_clear();
1606
1607 // Push all non-control nodes with no inputs from within block, then control entry
1608 for (int j = 0; j < _data_entry.length(); j++) {
1609 Node* n = _data_entry.at(j);
1610 visited_set(n);
1611 _stk.push(n);
1612 }
1613 visited_set(entry);
1614 _stk.push(entry);
1615
1616 // Do a depth first walk over out edges
1617 int rpo_idx = bb_ct - 1;
1618 int size;
1619 while ((size = _stk.length()) > 0) {
1620 Node* n = _stk.top(); // Leave node on stack
1621 if (!visited_test_set(n)) {
1622 // forward arc in graph
1623 } else if (!post_visited_test(n)) {
1624 // cross or back arc
1625 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1626 Node *use = n->fast_out(i);
1627 if (in_bb(use) && !visited_test(use) &&
1628 // Don't go around backedge
1629 (!use->is_Phi() || n == entry)) {
1630 _stk.push(use);
1631 }
1632 }
1633 if (_stk.length() == size) {
1634 // There were no additional uses, post visit node now
1635 _stk.pop(); // Remove node from stack
1636 assert(rpo_idx >= 0, "");
1637 _block.at_put_grow(rpo_idx, n);
1638 rpo_idx--;
1639 post_visited_set(n);
1640 assert(rpo_idx >= 0 || _stk.is_empty(), "");
1641 }
1642 } else {
1643 _stk.pop(); // Remove post-visited node from stack
1644 }
1645 }
1646
1647 // Create real map of block indices for nodes
1648 for (int j = 0; j < _block.length(); j++) {
1649 Node* n = _block.at(j);
1650 set_bb_idx(n, j);
1651 }
1652
1653 initialize_bb(); // Ensure extra info is allocated.
1654
1655 #ifndef PRODUCT
1656 if (TraceSuperWord) {
1657 print_bb();
1658 tty->print_cr("\ndata entry nodes: %s", _data_entry.length() > 0 ? "" : "NONE");
1659 for (int m = 0; m < _data_entry.length(); m++) {
1660 tty->print("%3d ", m);
1661 _data_entry.at(m)->dump();
1662 }
1663 tty->print_cr("\nmemory slices: %s", _mem_slice_head.length() > 0 ? "" : "NONE");
1664 for (int m = 0; m < _mem_slice_head.length(); m++) {
1665 tty->print("%3d ", m); _mem_slice_head.at(m)->dump();
1666 tty->print(" "); _mem_slice_tail.at(m)->dump();
1667 }
1668 }
1669 #endif
1670 assert(rpo_idx == -1 && bb_ct == _block.length(), "all block members found");
1671 }
1672
1673 //------------------------------initialize_bb---------------------------
1674 // Initialize per node info
1675 void SuperWord::initialize_bb() {
1676 Node* last = _block.at(_block.length() - 1);
1677 grow_node_info(bb_idx(last));
1678 }
1679
1680 //------------------------------bb_insert_after---------------------------
1681 // Insert n into block after pos
1682 void SuperWord::bb_insert_after(Node* n, int pos) {
1683 int n_pos = pos + 1;
1684 // Make room
1685 for (int i = _block.length() - 1; i >= n_pos; i--) {
1686 _block.at_put_grow(i+1, _block.at(i));
1687 }
1688 for (int j = _node_info.length() - 1; j >= n_pos; j--) {
1689 _node_info.at_put_grow(j+1, _node_info.at(j));
1690 }
1691 // Set value
1692 _block.at_put_grow(n_pos, n);
1693 _node_info.at_put_grow(n_pos, SWNodeInfo::initial);
1694 // Adjust map from node->_idx to _block index
1695 for (int i = n_pos; i < _block.length(); i++) {
1696 set_bb_idx(_block.at(i), i);
1697 }
1698 }
1699
1700 //------------------------------compute_max_depth---------------------------
1701 // Compute max depth for expressions from beginning of block
1702 // Use to prune search paths during test for independence.
1703 void SuperWord::compute_max_depth() {
1704 int ct = 0;
1705 bool again;
1706 do {
1707 again = false;
1708 for (int i = 0; i < _block.length(); i++) {
1709 Node* n = _block.at(i);
1710 if (!n->is_Phi()) {
1711 int d_orig = depth(n);
1712 int d_in = 0;
1713 for (DepPreds preds(n, _dg); !preds.done(); preds.next()) {
1714 Node* pred = preds.current();
1715 if (in_bb(pred)) {
1716 d_in = MAX2(d_in, depth(pred));
1717 }
1718 }
1719 if (d_in + 1 != d_orig) {
1720 set_depth(n, d_in + 1);
1721 again = true;
1722 }
1723 }
1724 }
1725 ct++;
1726 } while (again);
1727 #ifndef PRODUCT
1728 if (TraceSuperWord && Verbose)
1729 tty->print_cr("compute_max_depth iterated: %d times", ct);
1730 #endif
1731 }
1732
1733 //-------------------------compute_vector_element_type-----------------------
1734 // Compute necessary vector element type for expressions
1735 // This propagates backwards a narrower integer type when the
1736 // upper bits of the value are not needed.
1737 // Example: char a,b,c; a = b + c;
1738 // Normally the type of the add is integer, but for packed character
1739 // operations the type of the add needs to be char.
1740 void SuperWord::compute_vector_element_type() {
1741 #ifndef PRODUCT
1742 if (TraceSuperWord && Verbose)
1743 tty->print_cr("\ncompute_velt_type:");
1744 #endif
1745
1746 // Initial type
1747 for (int i = 0; i < _block.length(); i++) {
1748 Node* n = _block.at(i);
1749 set_velt_type(n, container_type(n));
1750 }
1751
1752 // Propagate narrowed type backwards through operations
1753 // that don't depend on higher order bits
1754 for (int i = _block.length() - 1; i >= 0; i--) {
1755 Node* n = _block.at(i);
1756 // Only integer types need be examined
1757 const Type* vt = velt_type(n);
1758 if (vt->basic_type() == T_INT) {
1759 uint start, end;
1760 vector_opd_range(n, &start, &end);
1761 const Type* vt = velt_type(n);
1762
1763 for (uint j = start; j < end; j++) {
1764 Node* in = n->in(j);
1765 // Don't propagate through a memory
1766 if (!in->is_Mem() && in_bb(in) && velt_type(in)->basic_type() == T_INT &&
1767 data_size(n) < data_size(in)) {
1768 bool same_type = true;
1769 for (DUIterator_Fast kmax, k = in->fast_outs(kmax); k < kmax; k++) {
1770 Node *use = in->fast_out(k);
1771 if (!in_bb(use) || !same_velt_type(use, n)) {
1772 same_type = false;
1773 break;
1774 }
1775 }
1776 if (same_type) {
1777 set_velt_type(in, vt);
1778 }
1779 }
1780 }
1781 }
1782 }
1783 #ifndef PRODUCT
1784 if (TraceSuperWord && Verbose) {
1785 for (int i = 0; i < _block.length(); i++) {
1786 Node* n = _block.at(i);
1787 velt_type(n)->dump();
1788 tty->print("\t");
1789 n->dump();
1790 }
1791 }
1792 #endif
1793 }
1794
1795 //------------------------------memory_alignment---------------------------
1796 // Alignment within a vector memory reference
1797 int SuperWord::memory_alignment(MemNode* s, int iv_adjust_in_bytes) {
1798 SWPointer p(s, this);
1799 if (!p.valid()) {
1800 return bottom_align;
1801 }
1802 int vw = vector_width_in_bytes(s);
1803 if (vw < 2) {
1804 return bottom_align; // No vectors for this type
1805 }
1806 int offset = p.offset_in_bytes();
1807 offset += iv_adjust_in_bytes;
1808 int off_rem = offset % vw;
1809 int off_mod = off_rem >= 0 ? off_rem : off_rem + vw;
1810 return off_mod;
1811 }
1812
1813 //---------------------------container_type---------------------------
1814 // Smallest type containing range of values
1815 const Type* SuperWord::container_type(Node* n) {
1816 if (n->is_Mem()) {
1817 return Type::get_const_basic_type(n->as_Mem()->memory_type());
1818 }
1819 const Type* t = _igvn.type(n);
1820 if (t->basic_type() == T_INT) {
1821 // A narrow type of arithmetic operations will be determined by
1822 // propagating the type of memory operations.
1823 return TypeInt::INT;
1824 }
1825 return t;
1826 }
1827
1828 bool SuperWord::same_velt_type(Node* n1, Node* n2) {
1829 const Type* vt1 = velt_type(n1);
1830 const Type* vt2 = velt_type(n1);
1831 if (vt1->basic_type() == T_INT && vt2->basic_type() == T_INT) {
1832 // Compare vectors element sizes for integer types.
1833 return data_size(n1) == data_size(n2);
1834 }
1835 return vt1 == vt2;
1836 }
1837
1838 //-------------------------vector_opd_range-----------------------
1839 // (Start, end] half-open range defining which operands are vector
1840 void SuperWord::vector_opd_range(Node* n, uint* start, uint* end) {
1841 switch (n->Opcode()) {
1842 case Op_LoadB: case Op_LoadUB:
1843 case Op_LoadS: case Op_LoadUS:
1844 case Op_LoadI: case Op_LoadL:
1845 case Op_LoadF: case Op_LoadD:
1846 case Op_LoadP:
1847 *start = 0;
1848 *end = 0;
1849 return;
1850 case Op_StoreB: case Op_StoreC:
1851 case Op_StoreI: case Op_StoreL:
1852 case Op_StoreF: case Op_StoreD:
1853 case Op_StoreP:
1854 *start = MemNode::ValueIn;
1855 *end = *start + 1;
1856 return;
1857 case Op_LShiftI: case Op_LShiftL:
1858 *start = 1;
1859 *end = 2;
1860 return;
1861 case Op_CMoveI: case Op_CMoveL: case Op_CMoveF: case Op_CMoveD:
1862 *start = 2;
1863 *end = n->req();
1864 return;
1865 }
1866 *start = 1;
1867 *end = n->req(); // default is all operands
1868 }
1869
1870 //------------------------------in_packset---------------------------
1871 // Are s1 and s2 in a pack pair and ordered as s1,s2?
1872 bool SuperWord::in_packset(Node* s1, Node* s2) {
1873 for (int i = 0; i < _packset.length(); i++) {
1874 Node_List* p = _packset.at(i);
1875 assert(p->size() == 2, "must be");
1876 if (p->at(0) == s1 && p->at(p->size()-1) == s2) {
1877 return true;
1878 }
1879 }
1880 return false;
1881 }
1882
1883 //------------------------------in_pack---------------------------
1884 // Is s in pack p?
1885 Node_List* SuperWord::in_pack(Node* s, Node_List* p) {
1886 for (uint i = 0; i < p->size(); i++) {
1887 if (p->at(i) == s) {
1888 return p;
1889 }
1890 }
1891 return NULL;
1892 }
1893
1894 //------------------------------remove_pack_at---------------------------
1895 // Remove the pack at position pos in the packset
1896 void SuperWord::remove_pack_at(int pos) {
1897 Node_List* p = _packset.at(pos);
1898 for (uint i = 0; i < p->size(); i++) {
1899 Node* s = p->at(i);
1900 set_my_pack(s, NULL);
1901 }
1902 _packset.remove_at(pos);
1903 }
1904
1905 //------------------------------executed_first---------------------------
1906 // Return the node executed first in pack p. Uses the RPO block list
1907 // to determine order.
1908 Node* SuperWord::executed_first(Node_List* p) {
1909 Node* n = p->at(0);
1910 int n_rpo = bb_idx(n);
1911 for (uint i = 1; i < p->size(); i++) {
1912 Node* s = p->at(i);
1913 int s_rpo = bb_idx(s);
1914 if (s_rpo < n_rpo) {
1915 n = s;
1916 n_rpo = s_rpo;
1917 }
1918 }
1919 return n;
1920 }
1921
1922 //------------------------------executed_last---------------------------
1923 // Return the node executed last in pack p.
1924 Node* SuperWord::executed_last(Node_List* p) {
1925 Node* n = p->at(0);
1926 int n_rpo = bb_idx(n);
1927 for (uint i = 1; i < p->size(); i++) {
1928 Node* s = p->at(i);
1929 int s_rpo = bb_idx(s);
1930 if (s_rpo > n_rpo) {
1931 n = s;
1932 n_rpo = s_rpo;
1933 }
1934 }
1935 return n;
1936 }
1937
1938 //----------------------------align_initial_loop_index---------------------------
1939 // Adjust pre-loop limit so that in main loop, a load/store reference
1940 // to align_to_ref will be a position zero in the vector.
1941 // (iv + k) mod vector_align == 0
1942 void SuperWord::align_initial_loop_index(MemNode* align_to_ref) {
1943 CountedLoopNode *main_head = lp()->as_CountedLoop();
1944 assert(main_head->is_main_loop(), "");
1945 CountedLoopEndNode* pre_end = get_pre_loop_end(main_head);
1946 assert(pre_end != NULL, "");
1947 Node *pre_opaq1 = pre_end->limit();
1948 assert(pre_opaq1->Opcode() == Op_Opaque1, "");
1949 Opaque1Node *pre_opaq = (Opaque1Node*)pre_opaq1;
1950 Node *lim0 = pre_opaq->in(1);
1951
1952 // Where we put new limit calculations
1953 Node *pre_ctrl = pre_end->loopnode()->in(LoopNode::EntryControl);
1954
1955 // Ensure the original loop limit is available from the
1956 // pre-loop Opaque1 node.
1957 Node *orig_limit = pre_opaq->original_loop_limit();
1958 assert(orig_limit != NULL && _igvn.type(orig_limit) != Type::TOP, "");
1959
1960 SWPointer align_to_ref_p(align_to_ref, this);
1961 assert(align_to_ref_p.valid(), "sanity");
1962
1963 // Given:
1964 // lim0 == original pre loop limit
1965 // V == v_align (power of 2)
1966 // invar == extra invariant piece of the address expression
1967 // e == k [ +/- invar ]
1968 //
1969 // When reassociating expressions involving '%' the basic rules are:
1970 // (a - b) % k == 0 => a % k == b % k
1971 // and:
1972 // (a + b) % k == 0 => a % k == (k - b) % k
1973 //
1974 // For stride > 0 && scale > 0,
1975 // Derive the new pre-loop limit "lim" such that the two constraints:
1976 // (1) lim = lim0 + N (where N is some positive integer < V)
1977 // (2) (e + lim) % V == 0
1978 // are true.
1979 //
1980 // Substituting (1) into (2),
1981 // (e + lim0 + N) % V == 0
1982 // solve for N:
1983 // N = (V - (e + lim0)) % V
1984 // substitute back into (1), so that new limit
1985 // lim = lim0 + (V - (e + lim0)) % V
1986 //
1987 // For stride > 0 && scale < 0
1988 // Constraints:
1989 // lim = lim0 + N
1990 // (e - lim) % V == 0
1991 // Solving for lim:
1992 // (e - lim0 - N) % V == 0
1993 // N = (e - lim0) % V
1994 // lim = lim0 + (e - lim0) % V
1995 //
1996 // For stride < 0 && scale > 0
1997 // Constraints:
1998 // lim = lim0 - N
1999 // (e + lim) % V == 0
2000 // Solving for lim:
2001 // (e + lim0 - N) % V == 0
2002 // N = (e + lim0) % V
2003 // lim = lim0 - (e + lim0) % V
2004 //
2005 // For stride < 0 && scale < 0
2006 // Constraints:
2007 // lim = lim0 - N
2008 // (e - lim) % V == 0
2009 // Solving for lim:
2010 // (e - lim0 + N) % V == 0
2011 // N = (V - (e - lim0)) % V
2012 // lim = lim0 - (V - (e - lim0)) % V
2013
2014 int vw = vector_width_in_bytes(align_to_ref);
2015 int stride = iv_stride();
2016 int scale = align_to_ref_p.scale_in_bytes();
2017 int elt_size = align_to_ref_p.memory_size();
2018 int v_align = vw / elt_size;
2019 assert(v_align > 1, "sanity");
2020 int k = align_to_ref_p.offset_in_bytes() / elt_size;
2021
2022 Node *kn = _igvn.intcon(k);
2023
2024 Node *e = kn;
2025 if (align_to_ref_p.invar() != NULL) {
2026 // incorporate any extra invariant piece producing k +/- invar >>> log2(elt)
2027 Node* log2_elt = _igvn.intcon(exact_log2(elt_size));
2028 Node* aref = new (_phase->C, 3) URShiftINode(align_to_ref_p.invar(), log2_elt);
2029 _phase->_igvn.register_new_node_with_optimizer(aref);
2030 _phase->set_ctrl(aref, pre_ctrl);
2031 if (align_to_ref_p.negate_invar()) {
2032 e = new (_phase->C, 3) SubINode(e, aref);
2033 } else {
2034 e = new (_phase->C, 3) AddINode(e, aref);
2035 }
2036 _phase->_igvn.register_new_node_with_optimizer(e);
2037 _phase->set_ctrl(e, pre_ctrl);
2038 }
2039 if (vw > ObjectAlignmentInBytes) {
2040 // incorporate base e +/- base && Mask >>> log2(elt)
2041 Node* mask = _igvn.MakeConX(~(-1 << exact_log2(vw)));
2042 Node* xbase = new(_phase->C, 2) CastP2XNode(NULL, align_to_ref_p.base());
2043 _phase->_igvn.register_new_node_with_optimizer(xbase);
2044 Node* masked_xbase = new (_phase->C, 3) AndXNode(xbase, mask);
2045 _phase->_igvn.register_new_node_with_optimizer(masked_xbase);
2046 #ifdef _LP64
2047 masked_xbase = new (_phase->C, 2) ConvL2INode(masked_xbase);
2048 _phase->_igvn.register_new_node_with_optimizer(masked_xbase);
2049 #endif
2050 Node* log2_elt = _igvn.intcon(exact_log2(elt_size));
2051 Node* bref = new (_phase->C, 3) URShiftINode(masked_xbase, log2_elt);
2052 _phase->_igvn.register_new_node_with_optimizer(bref);
2053 _phase->set_ctrl(bref, pre_ctrl);
2054 e = new (_phase->C, 3) AddINode(e, bref);
2055 _phase->_igvn.register_new_node_with_optimizer(e);
2056 _phase->set_ctrl(e, pre_ctrl);
2057 }
2058
2059 // compute e +/- lim0
2060 if (scale < 0) {
2061 e = new (_phase->C, 3) SubINode(e, lim0);
2062 } else {
2063 e = new (_phase->C, 3) AddINode(e, lim0);
2064 }
2065 _phase->_igvn.register_new_node_with_optimizer(e);
2066 _phase->set_ctrl(e, pre_ctrl);
2067
2068 if (stride * scale > 0) {
2069 // compute V - (e +/- lim0)
2070 Node* va = _igvn.intcon(v_align);
2071 e = new (_phase->C, 3) SubINode(va, e);
2072 _phase->_igvn.register_new_node_with_optimizer(e);
2073 _phase->set_ctrl(e, pre_ctrl);
2074 }
2075 // compute N = (exp) % V
2076 Node* va_msk = _igvn.intcon(v_align - 1);
2077 Node* N = new (_phase->C, 3) AndINode(e, va_msk);
2078 _phase->_igvn.register_new_node_with_optimizer(N);
2079 _phase->set_ctrl(N, pre_ctrl);
2080
2081 // substitute back into (1), so that new limit
2082 // lim = lim0 + N
2083 Node* lim;
2084 if (stride < 0) {
2085 lim = new (_phase->C, 3) SubINode(lim0, N);
2086 } else {
2087 lim = new (_phase->C, 3) AddINode(lim0, N);
2088 }
2089 _phase->_igvn.register_new_node_with_optimizer(lim);
2090 _phase->set_ctrl(lim, pre_ctrl);
2091 Node* constrained =
2092 (stride > 0) ? (Node*) new (_phase->C,3) MinINode(lim, orig_limit)
2093 : (Node*) new (_phase->C,3) MaxINode(lim, orig_limit);
2094 _phase->_igvn.register_new_node_with_optimizer(constrained);
2095 _phase->set_ctrl(constrained, pre_ctrl);
2096 _igvn.hash_delete(pre_opaq);
2097 pre_opaq->set_req(1, constrained);
2098 }
2099
2100 //----------------------------get_pre_loop_end---------------------------
2101 // Find pre loop end from main loop. Returns null if none.
2102 CountedLoopEndNode* SuperWord::get_pre_loop_end(CountedLoopNode *cl) {
2103 Node *ctrl = cl->in(LoopNode::EntryControl);
2104 if (!ctrl->is_IfTrue() && !ctrl->is_IfFalse()) return NULL;
2105 Node *iffm = ctrl->in(0);
2106 if (!iffm->is_If()) return NULL;
2107 Node *p_f = iffm->in(0);
2108 if (!p_f->is_IfFalse()) return NULL;
2109 if (!p_f->in(0)->is_CountedLoopEnd()) return NULL;
2110 CountedLoopEndNode *pre_end = p_f->in(0)->as_CountedLoopEnd();
2111 if (!pre_end->loopnode()->is_pre_loop()) return NULL;
2112 return pre_end;
2113 }
2114
2115
2116 //------------------------------init---------------------------
2117 void SuperWord::init() {
2118 _dg.init();
2119 _packset.clear();
2120 _disjoint_ptrs.clear();
2121 _block.clear();
2122 _data_entry.clear();
2123 _mem_slice_head.clear();
2124 _mem_slice_tail.clear();
2125 _node_info.clear();
2126 _align_to_ref = NULL;
2127 _lpt = NULL;
2128 _lp = NULL;
2129 _bb = NULL;
2130 _iv = NULL;
2131 }
2132
2133 //------------------------------print_packset---------------------------
2134 void SuperWord::print_packset() {
2135 #ifndef PRODUCT
2136 tty->print_cr("packset");
2137 for (int i = 0; i < _packset.length(); i++) {
2138 tty->print_cr("Pack: %d", i);
2139 Node_List* p = _packset.at(i);
2140 print_pack(p);
2141 }
2142 #endif
2143 }
2144
2145 //------------------------------print_pack---------------------------
2146 void SuperWord::print_pack(Node_List* p) {
2147 for (uint i = 0; i < p->size(); i++) {
2148 print_stmt(p->at(i));
2149 }
2150 }
2151
2152 //------------------------------print_bb---------------------------
2153 void SuperWord::print_bb() {
2154 #ifndef PRODUCT
2155 tty->print_cr("\nBlock");
2156 for (int i = 0; i < _block.length(); i++) {
2157 Node* n = _block.at(i);
2158 tty->print("%d ", i);
2159 if (n) {
2160 n->dump();
2161 }
2162 }
2163 #endif
2164 }
2165
2166 //------------------------------print_stmt---------------------------
2167 void SuperWord::print_stmt(Node* s) {
2168 #ifndef PRODUCT
2169 tty->print(" align: %d \t", alignment(s));
2170 s->dump();
2171 #endif
2172 }
2173
2174 //------------------------------blank---------------------------
2175 char* SuperWord::blank(uint depth) {
2176 static char blanks[101];
2177 assert(depth < 101, "too deep");
2178 for (uint i = 0; i < depth; i++) blanks[i] = ' ';
2179 blanks[depth] = '\0';
2180 return blanks;
2181 }
2182
2183
2184 //==============================SWPointer===========================
2185
2186 //----------------------------SWPointer------------------------
2187 SWPointer::SWPointer(MemNode* mem, SuperWord* slp) :
2188 _mem(mem), _slp(slp), _base(NULL), _adr(NULL),
2189 _scale(0), _offset(0), _invar(NULL), _negate_invar(false) {
2190
2191 Node* adr = mem->in(MemNode::Address);
2192 if (!adr->is_AddP()) {
2193 assert(!valid(), "too complex");
2194 return;
2195 }
2196 // Match AddP(base, AddP(ptr, k*iv [+ invariant]), constant)
2197 Node* base = adr->in(AddPNode::Base);
2198 //unsafe reference could not be aligned appropriately without runtime checking
2199 if (base == NULL || base->bottom_type() == Type::TOP) {
2200 assert(!valid(), "unsafe access");
2201 return;
2202 }
2203 for (int i = 0; i < 3; i++) {
2204 if (!scaled_iv_plus_offset(adr->in(AddPNode::Offset))) {
2205 assert(!valid(), "too complex");
2206 return;
2207 }
2208 adr = adr->in(AddPNode::Address);
2209 if (base == adr || !adr->is_AddP()) {
2210 break; // stop looking at addp's
2211 }
2212 }
2213 _base = base;
2214 _adr = adr;
2215 assert(valid(), "Usable");
2216 }
2217
2218 // Following is used to create a temporary object during
2219 // the pattern match of an address expression.
2220 SWPointer::SWPointer(SWPointer* p) :
2221 _mem(p->_mem), _slp(p->_slp), _base(NULL), _adr(NULL),
2222 _scale(0), _offset(0), _invar(NULL), _negate_invar(false) {}
2223
2224 //------------------------scaled_iv_plus_offset--------------------
2225 // Match: k*iv + offset
2226 // where: k is a constant that maybe zero, and
2227 // offset is (k2 [+/- invariant]) where k2 maybe zero and invariant is optional
2228 bool SWPointer::scaled_iv_plus_offset(Node* n) {
2229 if (scaled_iv(n)) {
2230 return true;
2231 }
2232 if (offset_plus_k(n)) {
2233 return true;
2234 }
2235 int opc = n->Opcode();
2236 if (opc == Op_AddI) {
2237 if (scaled_iv(n->in(1)) && offset_plus_k(n->in(2))) {
2238 return true;
2239 }
2240 if (scaled_iv(n->in(2)) && offset_plus_k(n->in(1))) {
2241 return true;
2242 }
2243 } else if (opc == Op_SubI) {
2244 if (scaled_iv(n->in(1)) && offset_plus_k(n->in(2), true)) {
2245 return true;
2246 }
2247 if (scaled_iv(n->in(2)) && offset_plus_k(n->in(1))) {
2248 _scale *= -1;
2249 return true;
2250 }
2251 }
2252 return false;
2253 }
2254
2255 //----------------------------scaled_iv------------------------
2256 // Match: k*iv where k is a constant that's not zero
2257 bool SWPointer::scaled_iv(Node* n) {
2258 if (_scale != 0) {
2259 return false; // already found a scale
2260 }
2261 if (n == iv()) {
2262 _scale = 1;
2263 return true;
2264 }
2265 int opc = n->Opcode();
2266 if (opc == Op_MulI) {
2267 if (n->in(1) == iv() && n->in(2)->is_Con()) {
2268 _scale = n->in(2)->get_int();
2269 return true;
2270 } else if (n->in(2) == iv() && n->in(1)->is_Con()) {
2271 _scale = n->in(1)->get_int();
2272 return true;
2273 }
2274 } else if (opc == Op_LShiftI) {
2275 if (n->in(1) == iv() && n->in(2)->is_Con()) {
2276 _scale = 1 << n->in(2)->get_int();
2277 return true;
2278 }
2279 } else if (opc == Op_ConvI2L) {
2280 if (scaled_iv_plus_offset(n->in(1))) {
2281 return true;
2282 }
2283 } else if (opc == Op_LShiftL) {
2284 if (!has_iv() && _invar == NULL) {
2285 // Need to preserve the current _offset value, so
2286 // create a temporary object for this expression subtree.
2287 // Hacky, so should re-engineer the address pattern match.
2288 SWPointer tmp(this);
2289 if (tmp.scaled_iv_plus_offset(n->in(1))) {
2290 if (tmp._invar == NULL) {
2291 int mult = 1 << n->in(2)->get_int();
2292 _scale = tmp._scale * mult;
2293 _offset += tmp._offset * mult;
2294 return true;
2295 }
2296 }
2297 }
2298 }
2299 return false;
2300 }
2301
2302 //----------------------------offset_plus_k------------------------
2303 // Match: offset is (k [+/- invariant])
2304 // where k maybe zero and invariant is optional, but not both.
2305 bool SWPointer::offset_plus_k(Node* n, bool negate) {
2306 int opc = n->Opcode();
2307 if (opc == Op_ConI) {
2308 _offset += negate ? -(n->get_int()) : n->get_int();
2309 return true;
2310 } else if (opc == Op_ConL) {
2311 // Okay if value fits into an int
2312 const TypeLong* t = n->find_long_type();
2313 if (t->higher_equal(TypeLong::INT)) {
2314 jlong loff = n->get_long();
2315 jint off = (jint)loff;
2316 _offset += negate ? -off : loff;
2317 return true;
2318 }
2319 return false;
2320 }
2321 if (_invar != NULL) return false; // already have an invariant
2322 if (opc == Op_AddI) {
2323 if (n->in(2)->is_Con() && invariant(n->in(1))) {
2324 _negate_invar = negate;
2325 _invar = n->in(1);
2326 _offset += negate ? -(n->in(2)->get_int()) : n->in(2)->get_int();
2327 return true;
2328 } else if (n->in(1)->is_Con() && invariant(n->in(2))) {
2329 _offset += negate ? -(n->in(1)->get_int()) : n->in(1)->get_int();
2330 _negate_invar = negate;
2331 _invar = n->in(2);
2332 return true;
2333 }
2334 }
2335 if (opc == Op_SubI) {
2336 if (n->in(2)->is_Con() && invariant(n->in(1))) {
2337 _negate_invar = negate;
2338 _invar = n->in(1);
2339 _offset += !negate ? -(n->in(2)->get_int()) : n->in(2)->get_int();
2340 return true;
2341 } else if (n->in(1)->is_Con() && invariant(n->in(2))) {
2342 _offset += negate ? -(n->in(1)->get_int()) : n->in(1)->get_int();
2343 _negate_invar = !negate;
2344 _invar = n->in(2);
2345 return true;
2346 }
2347 }
2348 if (invariant(n)) {
2349 _negate_invar = negate;
2350 _invar = n;
2351 return true;
2352 }
2353 return false;
2354 }
2355
2356 //----------------------------print------------------------
2357 void SWPointer::print() {
2358 #ifndef PRODUCT
2359 tty->print("base: %d adr: %d scale: %d offset: %d invar: %c%d\n",
2360 _base != NULL ? _base->_idx : 0,
2361 _adr != NULL ? _adr->_idx : 0,
2362 _scale, _offset,
2363 _negate_invar?'-':'+',
2364 _invar != NULL ? _invar->_idx : 0);
2365 #endif
2366 }
2367
2368 // ========================= OrderedPair =====================
2369
2370 const OrderedPair OrderedPair::initial;
2371
2372 // ========================= SWNodeInfo =====================
2373
2374 const SWNodeInfo SWNodeInfo::initial;
2375
2376
2377 // ============================ DepGraph ===========================
2378
2379 //------------------------------make_node---------------------------
2380 // Make a new dependence graph node for an ideal node.
2381 DepMem* DepGraph::make_node(Node* node) {
2382 DepMem* m = new (_arena) DepMem(node);
2383 if (node != NULL) {
2384 assert(_map.at_grow(node->_idx) == NULL, "one init only");
2385 _map.at_put_grow(node->_idx, m);
2386 }
2387 return m;
2388 }
2389
2390 //------------------------------make_edge---------------------------
2391 // Make a new dependence graph edge from dpred -> dsucc
2392 DepEdge* DepGraph::make_edge(DepMem* dpred, DepMem* dsucc) {
2393 DepEdge* e = new (_arena) DepEdge(dpred, dsucc, dsucc->in_head(), dpred->out_head());
2394 dpred->set_out_head(e);
2395 dsucc->set_in_head(e);
2396 return e;
2397 }
2398
2399 // ========================== DepMem ========================
2400
2401 //------------------------------in_cnt---------------------------
2402 int DepMem::in_cnt() {
2403 int ct = 0;
2404 for (DepEdge* e = _in_head; e != NULL; e = e->next_in()) ct++;
2405 return ct;
2406 }
2407
2408 //------------------------------out_cnt---------------------------
2409 int DepMem::out_cnt() {
2410 int ct = 0;
2411 for (DepEdge* e = _out_head; e != NULL; e = e->next_out()) ct++;
2412 return ct;
2413 }
2414
2415 //------------------------------print-----------------------------
2416 void DepMem::print() {
2417 #ifndef PRODUCT
2418 tty->print(" DepNode %d (", _node->_idx);
2419 for (DepEdge* p = _in_head; p != NULL; p = p->next_in()) {
2420 Node* pred = p->pred()->node();
2421 tty->print(" %d", pred != NULL ? pred->_idx : 0);
2422 }
2423 tty->print(") [");
2424 for (DepEdge* s = _out_head; s != NULL; s = s->next_out()) {
2425 Node* succ = s->succ()->node();
2426 tty->print(" %d", succ != NULL ? succ->_idx : 0);
2427 }
2428 tty->print_cr(" ]");
2429 #endif
2430 }
2431
2432 // =========================== DepEdge =========================
2433
2434 //------------------------------DepPreds---------------------------
2435 void DepEdge::print() {
2436 #ifndef PRODUCT
2437 tty->print_cr("DepEdge: %d [ %d ]", _pred->node()->_idx, _succ->node()->_idx);
2438 #endif
2439 }
2440
2441 // =========================== DepPreds =========================
2442 // Iterator over predecessor edges in the dependence graph.
2443
2444 //------------------------------DepPreds---------------------------
2445 DepPreds::DepPreds(Node* n, DepGraph& dg) {
2446 _n = n;
2447 _done = false;
2448 if (_n->is_Store() || _n->is_Load()) {
2449 _next_idx = MemNode::Address;
2450 _end_idx = n->req();
2451 _dep_next = dg.dep(_n)->in_head();
2452 } else if (_n->is_Mem()) {
2453 _next_idx = 0;
2454 _end_idx = 0;
2455 _dep_next = dg.dep(_n)->in_head();
2456 } else {
2457 _next_idx = 1;
2458 _end_idx = _n->req();
2459 _dep_next = NULL;
2460 }
2461 next();
2462 }
2463
2464 //------------------------------next---------------------------
2465 void DepPreds::next() {
2466 if (_dep_next != NULL) {
2467 _current = _dep_next->pred()->node();
2468 _dep_next = _dep_next->next_in();
2469 } else if (_next_idx < _end_idx) {
2470 _current = _n->in(_next_idx++);
2471 } else {
2472 _done = true;
2473 }
2474 }
2475
2476 // =========================== DepSuccs =========================
2477 // Iterator over successor edges in the dependence graph.
2478
2479 //------------------------------DepSuccs---------------------------
2480 DepSuccs::DepSuccs(Node* n, DepGraph& dg) {
2481 _n = n;
2482 _done = false;
2483 if (_n->is_Load()) {
2484 _next_idx = 0;
2485 _end_idx = _n->outcnt();
2486 _dep_next = dg.dep(_n)->out_head();
2487 } else if (_n->is_Mem() || _n->is_Phi() && _n->bottom_type() == Type::MEMORY) {
2488 _next_idx = 0;
2489 _end_idx = 0;
2490 _dep_next = dg.dep(_n)->out_head();
2491 } else {
2492 _next_idx = 0;
2493 _end_idx = _n->outcnt();
2494 _dep_next = NULL;
2495 }
2496 next();
2497 }
2498
2499 //-------------------------------next---------------------------
2500 void DepSuccs::next() {
2501 if (_dep_next != NULL) {
2502 _current = _dep_next->succ()->node();
2503 _dep_next = _dep_next->next_out();
2504 } else if (_next_idx < _end_idx) {
2505 _current = _n->raw_out(_next_idx++);
2506 } else {
2507 _done = true;
2508 }
2509 }