1 /*
2 * Copyright (c) 2007, 2015, 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/castnode.hpp"
31 #include "opto/convertnode.hpp"
32 #include "opto/divnode.hpp"
33 #include "opto/matcher.hpp"
34 #include "opto/memnode.hpp"
35 #include "opto/mulnode.hpp"
36 #include "opto/opcodes.hpp"
37 #include "opto/opaquenode.hpp"
38 #include "opto/superword.hpp"
39 #include "opto/vectornode.hpp"
40
41 //
42 // S U P E R W O R D T R A N S F O R M
43 //=============================================================================
44
45 //------------------------------SuperWord---------------------------
46 SuperWord::SuperWord(PhaseIdealLoop* phase) :
47 _phase(phase),
48 _igvn(phase->_igvn),
49 _arena(phase->C->comp_arena()),
50 _packset(arena(), 8, 0, NULL), // packs for the current block
51 _bb_idx(arena(), (int)(1.10 * phase->C->unique()), 0, 0), // node idx to index in bb
52 _block(arena(), 8, 0, NULL), // nodes in current block
53 _data_entry(arena(), 8, 0, NULL), // nodes with all inputs from outside
54 _mem_slice_head(arena(), 8, 0, NULL), // memory slice heads
55 _mem_slice_tail(arena(), 8, 0, NULL), // memory slice tails
56 _node_info(arena(), 8, 0, SWNodeInfo::initial), // info needed per node
57 _clone_map(phase->C->clone_map()), // map of nodes created in cloning
58 _align_to_ref(NULL), // memory reference to align vectors to
59 _disjoint_ptrs(arena(), 8, 0, OrderedPair::initial), // runtime disambiguated pointer pairs
60 _dg(_arena), // dependence graph
61 _visited(arena()), // visited node set
62 _post_visited(arena()), // post visited node set
63 _n_idx_list(arena(), 8), // scratch list of (node,index) pairs
64 _stk(arena(), 8, 0, NULL), // scratch stack of nodes
65 _nlist(arena(), 8, 0, NULL), // scratch list of nodes
66 _lpt(NULL), // loop tree node
67 _lp(NULL), // LoopNode
68 _bb(NULL), // basic block
69 _iv(NULL), // induction var
70 _race_possible(false), // cases where SDMU is true
71 _early_return(true), // analysis evaluations routine
72 _num_work_vecs(0), // amount of vector work we have
73 _num_reductions(0), // amount of reduction work we have
74 _do_vector_loop(phase->C->do_vector_loop()), // whether to do vectorization/simd style
75 _ii_first(-1), // first loop generation index - only if do_vector_loop()
76 _ii_last(-1), // last loop generation index - only if do_vector_loop()
77 _ii_order(arena(), 8, 0, 0),
78 _vector_loop_debug(phase->C->has_method() && phase->C->method_has_option("VectorizeDebug"))
79 {}
80
81 //------------------------------transform_loop---------------------------
82 void SuperWord::transform_loop(IdealLoopTree* lpt, bool do_optimization) {
83 assert(UseSuperWord, "should be");
84 // Do vectors exist on this architecture?
85 if (Matcher::vector_width_in_bytes(T_BYTE) < 2) return;
86
87 assert(lpt->_head->is_CountedLoop(), "must be");
88 CountedLoopNode *cl = lpt->_head->as_CountedLoop();
89
90 if (!cl->is_valid_counted_loop()) return; // skip malformed counted loop
91
92 if (!cl->is_main_loop() ) return; // skip normal, pre, and post loops
93
94 // Check for no control flow in body (other than exit)
95 Node *cl_exit = cl->loopexit();
96 if (cl_exit->in(0) != lpt->_head) return;
97
98 // Make sure the are no extra control users of the loop backedge
99 if (cl->back_control()->outcnt() != 1) {
100 return;
101 }
102
103 // Check for pre-loop ending with CountedLoopEnd(Bool(Cmp(x,Opaque1(limit))))
104 CountedLoopEndNode* pre_end = get_pre_loop_end(cl);
105 if (pre_end == NULL) return;
106 Node *pre_opaq1 = pre_end->limit();
107 if (pre_opaq1->Opcode() != Op_Opaque1) return;
108
109 init(); // initialize data structures
110
111 set_lpt(lpt);
112 set_lp(cl);
113
114 // For now, define one block which is the entire loop body
115 set_bb(cl);
116
117 if (do_optimization) {
118 assert(_packset.length() == 0, "packset must be empty");
119 SLP_extract();
120 }
121 }
122
123 //------------------------------early unrolling analysis------------------------------
124 void SuperWord::unrolling_analysis(CountedLoopNode *cl, int &local_loop_unroll_factor) {
125 bool not_slp = false;
126 ResourceMark rm;
127 size_t ignored_size = lpt()->_body.size();
128 int *ignored_loop_nodes = NEW_RESOURCE_ARRAY(int, ignored_size);
129 Node_Stack nstack((int)ignored_size);
130 Node *cl_exit = cl->loopexit();
131
132 // First clear the entries
133 for (uint i = 0; i < lpt()->_body.size(); i++) {
134 ignored_loop_nodes[i] = -1;
135 }
136
137 int max_vector = Matcher::max_vector_size(T_INT);
138
139 // Process the loop, some/all of the stack entries will not be in order, ergo
140 // need to preprocess the ignored initial state before we process the loop
141 for (uint i = 0; i < lpt()->_body.size(); i++) {
142 Node* n = lpt()->_body.at(i);
143 if (n == cl->incr() ||
144 n->is_reduction() ||
145 n->is_AddP() ||
146 n->is_Cmp() ||
147 n->is_IfTrue() ||
148 n->is_CountedLoop() ||
149 (n == cl_exit)) {
150 ignored_loop_nodes[i] = n->_idx;
151 continue;
152 }
153
154 if (n->is_If()) {
155 IfNode *iff = n->as_If();
156 if (iff->_fcnt != COUNT_UNKNOWN && iff->_prob != PROB_UNKNOWN) {
157 if (lpt()->is_loop_exit(iff)) {
158 ignored_loop_nodes[i] = n->_idx;
159 continue;
160 }
161 }
162 }
163
164 if (n->is_Phi() && (n->bottom_type() == Type::MEMORY)) {
165 Node* n_tail = n->in(LoopNode::LoopBackControl);
166 if (n_tail != n->in(LoopNode::EntryControl)) {
167 if (!n_tail->is_Mem()) {
168 not_slp = true;
169 break;
170 }
171 }
172 }
173
174 // This must happen after check of phi/if
175 if (n->is_Phi() || n->is_If()) {
176 ignored_loop_nodes[i] = n->_idx;
177 continue;
178 }
179
180 if (n->is_LoadStore() || n->is_MergeMem() ||
181 (n->is_Proj() && !n->as_Proj()->is_CFG())) {
182 not_slp = true;
183 break;
184 }
185
186 if (n->is_Mem()) {
187 Node* adr = n->in(MemNode::Address);
188 Node* n_ctrl = _phase->get_ctrl(adr);
189
190 // save a queue of post process nodes
191 if (n_ctrl != NULL && lpt()->is_member(_phase->get_loop(n_ctrl))) {
192 MemNode* current = n->as_Mem();
193 BasicType bt = current->memory_type();
194 if (is_java_primitive(bt) == false) {
195 ignored_loop_nodes[i] = n->_idx;
196 continue;
197 }
198
199 // Process the memory expression
200 int stack_idx = 0;
201 bool have_side_effects = true;
202 if (adr->is_AddP() == false) {
203 nstack.push(adr, stack_idx++);
204 }
205 else {
206 // Mark the components of the memory operation in nstack
207 SWPointer p1(current, this, &nstack, true);
208 have_side_effects = p1.node_stack()->is_nonempty();
209 }
210
211 // Process the pointer stack
212 while (have_side_effects) {
213 Node* pointer_node = nstack.node();
214 for (uint j = 0; j < lpt()->_body.size(); j++) {
215 Node* cur_node = lpt()->_body.at(j);
216 if (cur_node == pointer_node) {
217 ignored_loop_nodes[j] = cur_node->_idx;
218 break;
219 }
220 }
221 nstack.pop();
222 have_side_effects = nstack.is_nonempty();
223 }
224 }
225 }
226 }
227
228 if (not_slp == false) {
229 // Now we try to find the maximum supported consistent vector which the machine
230 // description can use
231 for (uint i = 0; i < lpt()->_body.size(); i++) {
232 if (ignored_loop_nodes[i] != -1) continue;
233
234 BasicType bt;
235 Node* n = lpt()->_body.at(i);
236 if (n->is_Store()) {
237 bt = n->as_Mem()->memory_type();
238 }
239 else {
240 bt = n->bottom_type()->basic_type();
241 }
242
243 int cur_max_vector = Matcher::max_vector_size(bt);
244
245 // If a max vector exists which is not larger than _local_loop_unroll_factor
246 // stop looking, we already have the max vector to map to.
247 if (cur_max_vector <= local_loop_unroll_factor) {
248 not_slp = true;
249 #ifndef PRODUCT
250 if (TraceSuperWordLoopUnrollAnalysis) {
251 tty->print_cr("slp analysis fails: unroll limit equals max vector\n");
252 }
253 #endif
254 break;
255 }
256
257 // Map the maximal common vector
258 if (VectorNode::implemented(n->Opcode(), cur_max_vector, bt)) {
259 if (cur_max_vector < max_vector) {
260 max_vector = cur_max_vector;
261 }
262 }
263 }
264 if (not_slp == false) {
265 local_loop_unroll_factor = max_vector;
266 }
267 cl->mark_passed_slp();
268 cl->set_slp_max_unroll(local_loop_unroll_factor);
269 }
270 }
271
272 //------------------------------SLP_extract---------------------------
273 // Extract the superword level parallelism
274 //
275 // 1) A reverse post-order of nodes in the block is constructed. By scanning
276 // this list from first to last, all definitions are visited before their uses.
277 //
278 // 2) A point-to-point dependence graph is constructed between memory references.
279 // This simplies the upcoming "independence" checker.
280 //
281 // 3) The maximum depth in the node graph from the beginning of the block
282 // to each node is computed. This is used to prune the graph search
283 // in the independence checker.
284 //
285 // 4) For integer types, the necessary bit width is propagated backwards
286 // from stores to allow packed operations on byte, char, and short
287 // integers. This reverses the promotion to type "int" that javac
288 // did for operations like: char c1,c2,c3; c1 = c2 + c3.
289 //
290 // 5) One of the memory references is picked to be an aligned vector reference.
291 // The pre-loop trip count is adjusted to align this reference in the
292 // unrolled body.
293 //
294 // 6) The initial set of pack pairs is seeded with memory references.
295 //
296 // 7) The set of pack pairs is extended by following use->def and def->use links.
297 //
298 // 8) The pairs are combined into vector sized packs.
299 //
300 // 9) Reorder the memory slices to co-locate members of the memory packs.
301 //
302 // 10) Generate ideal vector nodes for the final set of packs and where necessary,
303 // inserting scalar promotion, vector creation from multiple scalars, and
304 // extraction of scalar values from vectors.
305 //
306 void SuperWord::SLP_extract() {
307
308 #ifndef PRODUCT
309 if (_do_vector_loop && TraceSuperWord) {
310 tty->print("SuperWord::SLP_extract\n");
311 tty->print("input loop\n");
312 _lpt->dump_head();
313 _lpt->dump();
314 for (uint i = 0; i < _lpt->_body.size(); i++) {
315 _lpt->_body.at(i)->dump();
316 }
317 }
318 #endif
319 // Ready the block
320 if (!construct_bb()) {
321 return; // Exit if no interesting nodes or complex graph.
322 }
323 // build _dg, _disjoint_ptrs
324 dependence_graph();
325
326 // compute function depth(Node*)
327 compute_max_depth();
328
329 if (_do_vector_loop) {
330 if (mark_generations() != -1) {
331 hoist_loads_in_graph(); // this only rebuild the graph; all basic structs need rebuild explicitly
332
333 if (!construct_bb()) {
334 return; // Exit if no interesting nodes or complex graph.
335 }
336 dependence_graph();
337 compute_max_depth();
338 }
339
340 #ifndef PRODUCT
341 if (TraceSuperWord) {
342 tty->print_cr("\nSuperWord::_do_vector_loop: graph after hoist_loads_in_graph");
343 _lpt->dump_head();
344 for (int j = 0; j < _block.length(); j++) {
345 Node* n = _block.at(j);
346 int d = depth(n);
347 for (int i = 0; i < d; i++) tty->print("%s", " ");
348 tty->print("%d :", d);
349 n->dump();
350 }
351 }
352 #endif
353 }
354
355 compute_vector_element_type();
356
357 // Attempt vectorization
358
359 find_adjacent_refs();
360
361 extend_packlist();
362
363 if (_do_vector_loop) {
364 if (_packset.length() == 0) {
365 #ifndef PRODUCT
366 if (TraceSuperWord) {
367 tty->print_cr("\nSuperWord::_do_vector_loop DFA could not build packset, now trying to build anyway");
368 }
369 #endif
370 pack_parallel();
371 }
372 }
373
374 combine_packs();
375
376 construct_my_pack_map();
377
378 filter_packs();
379
380 schedule();
381
382 output();
383 }
384
385 //------------------------------find_adjacent_refs---------------------------
386 // Find the adjacent memory references and create pack pairs for them.
387 // This is the initial set of packs that will then be extended by
388 // following use->def and def->use links. The align positions are
389 // assigned relative to the reference "align_to_ref"
390 void SuperWord::find_adjacent_refs() {
391 // Get list of memory operations
392 Node_List memops;
393 for (int i = 0; i < _block.length(); i++) {
394 Node* n = _block.at(i);
395 if (n->is_Mem() && !n->is_LoadStore() && in_bb(n) &&
396 is_java_primitive(n->as_Mem()->memory_type())) {
397 int align = memory_alignment(n->as_Mem(), 0);
398 if (align != bottom_align) {
399 memops.push(n);
400 }
401 }
402 }
403
404 Node_List align_to_refs;
405 int best_iv_adjustment = 0;
406 MemNode* best_align_to_mem_ref = NULL;
407
408 while (memops.size() != 0) {
409 // Find a memory reference to align to.
410 MemNode* mem_ref = find_align_to_ref(memops);
411 if (mem_ref == NULL) break;
412 align_to_refs.push(mem_ref);
413 int iv_adjustment = get_iv_adjustment(mem_ref);
414
415 if (best_align_to_mem_ref == NULL) {
416 // Set memory reference which is the best from all memory operations
417 // to be used for alignment. The pre-loop trip count is modified to align
418 // this reference to a vector-aligned address.
419 best_align_to_mem_ref = mem_ref;
420 best_iv_adjustment = iv_adjustment;
421 }
422
423 SWPointer align_to_ref_p(mem_ref, this, NULL, false);
424 // Set alignment relative to "align_to_ref" for all related memory operations.
425 for (int i = memops.size() - 1; i >= 0; i--) {
426 MemNode* s = memops.at(i)->as_Mem();
427 if (isomorphic(s, mem_ref)) {
428 SWPointer p2(s, this, NULL, false);
429 if (p2.comparable(align_to_ref_p)) {
430 int align = memory_alignment(s, iv_adjustment);
431 set_alignment(s, align);
432 }
433 }
434 }
435
436 // Create initial pack pairs of memory operations for which
437 // alignment is set and vectors will be aligned.
438 bool create_pack = true;
439 if (memory_alignment(mem_ref, best_iv_adjustment) == 0 || _do_vector_loop) {
440 if (!Matcher::misaligned_vectors_ok()) {
441 int vw = vector_width(mem_ref);
442 int vw_best = vector_width(best_align_to_mem_ref);
443 if (vw > vw_best) {
444 // Do not vectorize a memory access with more elements per vector
445 // if unaligned memory access is not allowed because number of
446 // iterations in pre-loop will be not enough to align it.
447 create_pack = false;
448 } else {
449 SWPointer p2(best_align_to_mem_ref, this, NULL, false);
450 if (align_to_ref_p.invar() != p2.invar()) {
451 // Do not vectorize memory accesses with different invariants
452 // if unaligned memory accesses are not allowed.
453 create_pack = false;
454 }
455 }
456 }
457 } else {
458 if (same_velt_type(mem_ref, best_align_to_mem_ref)) {
459 // Can't allow vectorization of unaligned memory accesses with the
460 // same type since it could be overlapped accesses to the same array.
461 create_pack = false;
462 } else {
463 // Allow independent (different type) unaligned memory operations
464 // if HW supports them.
465 if (!Matcher::misaligned_vectors_ok()) {
466 create_pack = false;
467 } else {
468 // Check if packs of the same memory type but
469 // with a different alignment were created before.
470 for (uint i = 0; i < align_to_refs.size(); i++) {
471 MemNode* mr = align_to_refs.at(i)->as_Mem();
472 if (same_velt_type(mr, mem_ref) &&
473 memory_alignment(mr, iv_adjustment) != 0)
474 create_pack = false;
475 }
476 }
477 }
478 }
479 if (create_pack) {
480 for (uint i = 0; i < memops.size(); i++) {
481 Node* s1 = memops.at(i);
482 int align = alignment(s1);
483 if (align == top_align) continue;
484 for (uint j = 0; j < memops.size(); j++) {
485 Node* s2 = memops.at(j);
486 if (alignment(s2) == top_align) continue;
487 if (s1 != s2 && are_adjacent_refs(s1, s2)) {
488 if (stmts_can_pack(s1, s2, align)) {
489 Node_List* pair = new Node_List();
490 pair->push(s1);
491 pair->push(s2);
492 if (!_do_vector_loop || _clone_map.idx(s1->_idx) == _clone_map.idx(s2->_idx)) {
493 _packset.append(pair);
494 }
495 }
496 }
497 }
498 }
499 } else { // Don't create unaligned pack
500 // First, remove remaining memory ops of the same type from the list.
501 for (int i = memops.size() - 1; i >= 0; i--) {
502 MemNode* s = memops.at(i)->as_Mem();
503 if (same_velt_type(s, mem_ref)) {
504 memops.remove(i);
505 }
506 }
507
508 // Second, remove already constructed packs of the same type.
509 for (int i = _packset.length() - 1; i >= 0; i--) {
510 Node_List* p = _packset.at(i);
511 MemNode* s = p->at(0)->as_Mem();
512 if (same_velt_type(s, mem_ref)) {
513 remove_pack_at(i);
514 }
515 }
516
517 // If needed find the best memory reference for loop alignment again.
518 if (same_velt_type(mem_ref, best_align_to_mem_ref)) {
519 // Put memory ops from remaining packs back on memops list for
520 // the best alignment search.
521 uint orig_msize = memops.size();
522 for (int i = 0; i < _packset.length(); i++) {
523 Node_List* p = _packset.at(i);
524 MemNode* s = p->at(0)->as_Mem();
525 assert(!same_velt_type(s, mem_ref), "sanity");
526 memops.push(s);
527 }
528 MemNode* best_align_to_mem_ref = find_align_to_ref(memops);
529 if (best_align_to_mem_ref == NULL) break;
530 best_iv_adjustment = get_iv_adjustment(best_align_to_mem_ref);
531 // Restore list.
532 while (memops.size() > orig_msize)
533 (void)memops.pop();
534 }
535 } // unaligned memory accesses
536
537 // Remove used mem nodes.
538 for (int i = memops.size() - 1; i >= 0; i--) {
539 MemNode* m = memops.at(i)->as_Mem();
540 if (alignment(m) != top_align) {
541 memops.remove(i);
542 }
543 }
544
545 } // while (memops.size() != 0
546 set_align_to_ref(best_align_to_mem_ref);
547
548 #ifndef PRODUCT
549 if (TraceSuperWord) {
550 tty->print_cr("\nAfter find_adjacent_refs");
551 print_packset();
552 }
553 #endif
554 }
555
556 //------------------------------find_align_to_ref---------------------------
557 // Find a memory reference to align the loop induction variable to.
558 // Looks first at stores then at loads, looking for a memory reference
559 // with the largest number of references similar to it.
560 MemNode* SuperWord::find_align_to_ref(Node_List &memops) {
561 GrowableArray<int> cmp_ct(arena(), memops.size(), memops.size(), 0);
562
563 // Count number of comparable memory ops
564 for (uint i = 0; i < memops.size(); i++) {
565 MemNode* s1 = memops.at(i)->as_Mem();
566 SWPointer p1(s1, this, NULL, false);
567 // Discard if pre loop can't align this reference
568 if (!ref_is_alignable(p1)) {
569 *cmp_ct.adr_at(i) = 0;
570 continue;
571 }
572 for (uint j = i+1; j < memops.size(); j++) {
573 MemNode* s2 = memops.at(j)->as_Mem();
574 if (isomorphic(s1, s2)) {
575 SWPointer p2(s2, this, NULL, false);
576 if (p1.comparable(p2)) {
577 (*cmp_ct.adr_at(i))++;
578 (*cmp_ct.adr_at(j))++;
579 }
580 }
581 }
582 }
583
584 // Find Store (or Load) with the greatest number of "comparable" references,
585 // biggest vector size, smallest data size and smallest iv offset.
586 int max_ct = 0;
587 int max_vw = 0;
588 int max_idx = -1;
589 int min_size = max_jint;
590 int min_iv_offset = max_jint;
591 for (uint j = 0; j < memops.size(); j++) {
592 MemNode* s = memops.at(j)->as_Mem();
593 if (s->is_Store()) {
594 int vw = vector_width_in_bytes(s);
595 assert(vw > 1, "sanity");
596 SWPointer p(s, this, NULL, false);
597 if (cmp_ct.at(j) > max_ct ||
598 cmp_ct.at(j) == max_ct &&
599 (vw > max_vw ||
600 vw == max_vw &&
601 (data_size(s) < min_size ||
602 data_size(s) == min_size &&
603 (p.offset_in_bytes() < min_iv_offset)))) {
604 max_ct = cmp_ct.at(j);
605 max_vw = vw;
606 max_idx = j;
607 min_size = data_size(s);
608 min_iv_offset = p.offset_in_bytes();
609 }
610 }
611 }
612 // If no stores, look at loads
613 if (max_ct == 0) {
614 for (uint j = 0; j < memops.size(); j++) {
615 MemNode* s = memops.at(j)->as_Mem();
616 if (s->is_Load()) {
617 int vw = vector_width_in_bytes(s);
618 assert(vw > 1, "sanity");
619 SWPointer p(s, this, NULL, false);
620 if (cmp_ct.at(j) > max_ct ||
621 cmp_ct.at(j) == max_ct &&
622 (vw > max_vw ||
623 vw == max_vw &&
624 (data_size(s) < min_size ||
625 data_size(s) == min_size &&
626 (p.offset_in_bytes() < min_iv_offset)))) {
627 max_ct = cmp_ct.at(j);
628 max_vw = vw;
629 max_idx = j;
630 min_size = data_size(s);
631 min_iv_offset = p.offset_in_bytes();
632 }
633 }
634 }
635 }
636
637 #ifdef ASSERT
638 if (TraceSuperWord && Verbose) {
639 tty->print_cr("\nVector memops after find_align_to_ref");
640 for (uint i = 0; i < memops.size(); i++) {
641 MemNode* s = memops.at(i)->as_Mem();
642 s->dump();
643 }
644 }
645 #endif
646
647 if (max_ct > 0) {
648 #ifdef ASSERT
649 if (TraceSuperWord) {
650 tty->print("\nVector align to node: ");
651 memops.at(max_idx)->as_Mem()->dump();
652 }
653 #endif
654 return memops.at(max_idx)->as_Mem();
655 }
656 return NULL;
657 }
658
659 //------------------------------ref_is_alignable---------------------------
660 // Can the preloop align the reference to position zero in the vector?
661 bool SuperWord::ref_is_alignable(SWPointer& p) {
662 if (!p.has_iv()) {
663 return true; // no induction variable
664 }
665 CountedLoopEndNode* pre_end = get_pre_loop_end(lp()->as_CountedLoop());
666 assert(pre_end != NULL, "we must have a correct pre-loop");
667 assert(pre_end->stride_is_con(), "pre loop stride is constant");
668 int preloop_stride = pre_end->stride_con();
669
670 int span = preloop_stride * p.scale_in_bytes();
671 int mem_size = p.memory_size();
672 int offset = p.offset_in_bytes();
673 // Stride one accesses are alignable if offset is aligned to memory operation size.
674 // Offset can be unaligned when UseUnalignedAccesses is used.
675 if (ABS(span) == mem_size && (ABS(offset) % mem_size) == 0) {
676 return true;
677 }
678 // If the initial offset from start of the object is computable,
679 // check if the pre-loop can align the final offset accordingly.
680 //
681 // In other words: Can we find an i such that the offset
682 // after i pre-loop iterations is aligned to vw?
683 // (init_offset + pre_loop) % vw == 0 (1)
684 // where
685 // pre_loop = i * span
686 // is the number of bytes added to the offset by i pre-loop iterations.
687 //
688 // For this to hold we need pre_loop to increase init_offset by
689 // pre_loop = vw - (init_offset % vw)
690 //
691 // This is only possible if pre_loop is divisible by span because each
692 // pre-loop iteration increases the initial offset by 'span' bytes:
693 // (vw - (init_offset % vw)) % span == 0
694 //
695 int vw = vector_width_in_bytes(p.mem());
696 assert(vw > 1, "sanity");
697 Node* init_nd = pre_end->init_trip();
698 if (init_nd->is_Con() && p.invar() == NULL) {
699 int init = init_nd->bottom_type()->is_int()->get_con();
700 int init_offset = init * p.scale_in_bytes() + offset;
701 assert(init_offset >= 0, "positive offset from object start");
702 if (vw % span == 0) {
703 // If vm is a multiple of span, we use formula (1).
704 if (span > 0) {
705 return (vw - (init_offset % vw)) % span == 0;
706 } else {
707 assert(span < 0, "nonzero stride * scale");
708 return (init_offset % vw) % -span == 0;
709 }
710 } else if (span % vw == 0) {
711 // If span is a multiple of vw, we can simplify formula (1) to:
712 // (init_offset + i * span) % vw == 0
713 // =>
714 // (init_offset % vw) + ((i * span) % vw) == 0
715 // =>
716 // init_offset % vw == 0
717 //
718 // Because we add a multiple of vw to the initial offset, the final
719 // offset is a multiple of vw if and only if init_offset is a multiple.
720 //
721 return (init_offset % vw) == 0;
722 }
723 }
724 return false;
725 }
726
727 //---------------------------get_iv_adjustment---------------------------
728 // Calculate loop's iv adjustment for this memory ops.
729 int SuperWord::get_iv_adjustment(MemNode* mem_ref) {
730 SWPointer align_to_ref_p(mem_ref, this, NULL, false);
731 int offset = align_to_ref_p.offset_in_bytes();
732 int scale = align_to_ref_p.scale_in_bytes();
733 int elt_size = align_to_ref_p.memory_size();
734 int vw = vector_width_in_bytes(mem_ref);
735 assert(vw > 1, "sanity");
736 int iv_adjustment;
737 if (scale != 0) {
738 int stride_sign = (scale * iv_stride()) > 0 ? 1 : -1;
739 // At least one iteration is executed in pre-loop by default. As result
740 // several iterations are needed to align memory operations in main-loop even
741 // if offset is 0.
742 int iv_adjustment_in_bytes = (stride_sign * vw - (offset % vw));
743 assert(((ABS(iv_adjustment_in_bytes) % elt_size) == 0),
744 err_msg_res("(%d) should be divisible by (%d)", iv_adjustment_in_bytes, elt_size));
745 iv_adjustment = iv_adjustment_in_bytes/elt_size;
746 } else {
747 // This memory op is not dependent on iv (scale == 0)
748 iv_adjustment = 0;
749 }
750
751 #ifndef PRODUCT
752 if (TraceSuperWord)
753 tty->print_cr("\noffset = %d iv_adjust = %d elt_size = %d scale = %d iv_stride = %d vect_size %d",
754 offset, iv_adjustment, elt_size, scale, iv_stride(), vw);
755 #endif
756 return iv_adjustment;
757 }
758
759 //---------------------------dependence_graph---------------------------
760 // Construct dependency graph.
761 // Add dependence edges to load/store nodes for memory dependence
762 // A.out()->DependNode.in(1) and DependNode.out()->B.prec(x)
763 void SuperWord::dependence_graph() {
764 // First, assign a dependence node to each memory node
765 for (int i = 0; i < _block.length(); i++ ) {
766 Node *n = _block.at(i);
767 if (n->is_Mem() || n->is_Phi() && n->bottom_type() == Type::MEMORY) {
768 _dg.make_node(n);
769 }
770 }
771
772 // For each memory slice, create the dependences
773 for (int i = 0; i < _mem_slice_head.length(); i++) {
774 Node* n = _mem_slice_head.at(i);
775 Node* n_tail = _mem_slice_tail.at(i);
776
777 // Get slice in predecessor order (last is first)
778 mem_slice_preds(n_tail, n, _nlist);
779
780 #ifndef PRODUCT
781 if(TraceSuperWord && Verbose) {
782 tty->print_cr("SuperWord::dependence_graph: built a new mem slice");
783 for (int j = _nlist.length() - 1; j >= 0 ; j--) {
784 _nlist.at(j)->dump();
785 }
786 }
787 #endif
788 // Make the slice dependent on the root
789 DepMem* slice = _dg.dep(n);
790 _dg.make_edge(_dg.root(), slice);
791
792 // Create a sink for the slice
793 DepMem* slice_sink = _dg.make_node(NULL);
794 _dg.make_edge(slice_sink, _dg.tail());
795
796 // Now visit each pair of memory ops, creating the edges
797 for (int j = _nlist.length() - 1; j >= 0 ; j--) {
798 Node* s1 = _nlist.at(j);
799
800 // If no dependency yet, use slice
801 if (_dg.dep(s1)->in_cnt() == 0) {
802 _dg.make_edge(slice, s1);
803 }
804 SWPointer p1(s1->as_Mem(), this, NULL, false);
805 bool sink_dependent = true;
806 for (int k = j - 1; k >= 0; k--) {
807 Node* s2 = _nlist.at(k);
808 if (s1->is_Load() && s2->is_Load())
809 continue;
810 SWPointer p2(s2->as_Mem(), this, NULL, false);
811
812 int cmp = p1.cmp(p2);
813 if (SuperWordRTDepCheck &&
814 p1.base() != p2.base() && p1.valid() && p2.valid()) {
815 // Create a runtime check to disambiguate
816 OrderedPair pp(p1.base(), p2.base());
817 _disjoint_ptrs.append_if_missing(pp);
818 } else if (!SWPointer::not_equal(cmp)) {
819 // Possibly same address
820 _dg.make_edge(s1, s2);
821 sink_dependent = false;
822 }
823 }
824 if (sink_dependent) {
825 _dg.make_edge(s1, slice_sink);
826 }
827 }
828 #ifndef PRODUCT
829 if (TraceSuperWord) {
830 tty->print_cr("\nDependence graph for slice: %d", n->_idx);
831 for (int q = 0; q < _nlist.length(); q++) {
832 _dg.print(_nlist.at(q));
833 }
834 tty->cr();
835 }
836 #endif
837 _nlist.clear();
838 }
839
840 #ifndef PRODUCT
841 if (TraceSuperWord) {
842 tty->print_cr("\ndisjoint_ptrs: %s", _disjoint_ptrs.length() > 0 ? "" : "NONE");
843 for (int r = 0; r < _disjoint_ptrs.length(); r++) {
844 _disjoint_ptrs.at(r).print();
845 tty->cr();
846 }
847 tty->cr();
848 }
849 #endif
850 }
851
852 //---------------------------mem_slice_preds---------------------------
853 // Return a memory slice (node list) in predecessor order starting at "start"
854 void SuperWord::mem_slice_preds(Node* start, Node* stop, GrowableArray<Node*> &preds) {
855 assert(preds.length() == 0, "start empty");
856 Node* n = start;
857 Node* prev = NULL;
858 while (true) {
859 assert(in_bb(n), "must be in block");
860 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
861 Node* out = n->fast_out(i);
862 if (out->is_Load()) {
863 if (in_bb(out)) {
864 preds.push(out);
865 }
866 } else {
867 // FIXME
868 if (out->is_MergeMem() && !in_bb(out)) {
869 // Either unrolling is causing a memory edge not to disappear,
870 // or need to run igvn.optimize() again before SLP
871 } else if (out->is_Phi() && out->bottom_type() == Type::MEMORY && !in_bb(out)) {
872 // Ditto. Not sure what else to check further.
873 } else if (out->Opcode() == Op_StoreCM && out->in(MemNode::OopStore) == n) {
874 // StoreCM has an input edge used as a precedence edge.
875 // Maybe an issue when oop stores are vectorized.
876 } else {
877 assert(out == prev || prev == NULL, "no branches off of store slice");
878 }
879 }
880 }
881 if (n == stop) break;
882 preds.push(n);
883 prev = n;
884 assert(n->is_Mem(), err_msg_res("unexpected node %s", n->Name()));
885 n = n->in(MemNode::Memory);
886 }
887 }
888
889 //------------------------------stmts_can_pack---------------------------
890 // Can s1 and s2 be in a pack with s1 immediately preceding s2 and
891 // s1 aligned at "align"
892 bool SuperWord::stmts_can_pack(Node* s1, Node* s2, int align) {
893
894 // Do not use superword for non-primitives
895 BasicType bt1 = velt_basic_type(s1);
896 BasicType bt2 = velt_basic_type(s2);
897 if(!is_java_primitive(bt1) || !is_java_primitive(bt2))
898 return false;
899 if (Matcher::max_vector_size(bt1) < 2) {
900 return false; // No vectors for this type
901 }
902
903 if (isomorphic(s1, s2)) {
904 if (independent(s1, s2) || reduction(s1, s2)) {
905 if (!exists_at(s1, 0) && !exists_at(s2, 1)) {
906 if (!s1->is_Mem() || are_adjacent_refs(s1, s2)) {
907 int s1_align = alignment(s1);
908 int s2_align = alignment(s2);
909 if (s1_align == top_align || s1_align == align) {
910 if (s2_align == top_align || s2_align == align + data_size(s1)) {
911 return true;
912 }
913 }
914 }
915 }
916 }
917 }
918 return false;
919 }
920
921 //------------------------------exists_at---------------------------
922 // Does s exist in a pack at position pos?
923 bool SuperWord::exists_at(Node* s, uint pos) {
924 for (int i = 0; i < _packset.length(); i++) {
925 Node_List* p = _packset.at(i);
926 if (p->at(pos) == s) {
927 return true;
928 }
929 }
930 return false;
931 }
932
933 //------------------------------are_adjacent_refs---------------------------
934 // Is s1 immediately before s2 in memory?
935 bool SuperWord::are_adjacent_refs(Node* s1, Node* s2) {
936 if (!s1->is_Mem() || !s2->is_Mem()) return false;
937 if (!in_bb(s1) || !in_bb(s2)) return false;
938
939 // Do not use superword for non-primitives
940 if (!is_java_primitive(s1->as_Mem()->memory_type()) ||
941 !is_java_primitive(s2->as_Mem()->memory_type())) {
942 return false;
943 }
944
945 // FIXME - co_locate_pack fails on Stores in different mem-slices, so
946 // only pack memops that are in the same alias set until that's fixed.
947 if (_phase->C->get_alias_index(s1->as_Mem()->adr_type()) !=
948 _phase->C->get_alias_index(s2->as_Mem()->adr_type()))
949 return false;
950 SWPointer p1(s1->as_Mem(), this, NULL, false);
951 SWPointer p2(s2->as_Mem(), this, NULL, false);
952 if (p1.base() != p2.base() || !p1.comparable(p2)) return false;
953 int diff = p2.offset_in_bytes() - p1.offset_in_bytes();
954 return diff == data_size(s1);
955 }
956
957 //------------------------------isomorphic---------------------------
958 // Are s1 and s2 similar?
959 bool SuperWord::isomorphic(Node* s1, Node* s2) {
960 if (s1->Opcode() != s2->Opcode()) return false;
961 if (s1->req() != s2->req()) return false;
962 if (s1->in(0) != s2->in(0)) return false;
963 if (!same_velt_type(s1, s2)) return false;
964 return true;
965 }
966
967 //------------------------------independent---------------------------
968 // Is there no data path from s1 to s2 or s2 to s1?
969 bool SuperWord::independent(Node* s1, Node* s2) {
970 // assert(s1->Opcode() == s2->Opcode(), "check isomorphic first");
971 int d1 = depth(s1);
972 int d2 = depth(s2);
973 if (d1 == d2) return s1 != s2;
974 Node* deep = d1 > d2 ? s1 : s2;
975 Node* shallow = d1 > d2 ? s2 : s1;
976
977 visited_clear();
978
979 return independent_path(shallow, deep);
980 }
981
982 //------------------------------reduction---------------------------
983 // Is there a data path between s1 and s2 and the nodes reductions?
984 bool SuperWord::reduction(Node* s1, Node* s2) {
985 bool retValue = false;
986 int d1 = depth(s1);
987 int d2 = depth(s2);
988 if (d1 + 1 == d2) {
989 if (s1->is_reduction() && s2->is_reduction()) {
990 // This is an ordered set, so s1 should define s2
991 for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
992 Node* t1 = s1->fast_out(i);
993 if (t1 == s2) {
994 // both nodes are reductions and connected
995 retValue = true;
996 }
997 }
998 }
999 }
1000
1001 return retValue;
1002 }
1003
1004 //------------------------------independent_path------------------------------
1005 // Helper for independent
1006 bool SuperWord::independent_path(Node* shallow, Node* deep, uint dp) {
1007 if (dp >= 1000) return false; // stop deep recursion
1008 visited_set(deep);
1009 int shal_depth = depth(shallow);
1010 assert(shal_depth <= depth(deep), "must be");
1011 for (DepPreds preds(deep, _dg); !preds.done(); preds.next()) {
1012 Node* pred = preds.current();
1013 if (in_bb(pred) && !visited_test(pred)) {
1014 if (shallow == pred) {
1015 return false;
1016 }
1017 if (shal_depth < depth(pred) && !independent_path(shallow, pred, dp+1)) {
1018 return false;
1019 }
1020 }
1021 }
1022 return true;
1023 }
1024
1025 //------------------------------set_alignment---------------------------
1026 void SuperWord::set_alignment(Node* s1, Node* s2, int align) {
1027 set_alignment(s1, align);
1028 if (align == top_align || align == bottom_align) {
1029 set_alignment(s2, align);
1030 } else {
1031 set_alignment(s2, align + data_size(s1));
1032 }
1033 }
1034
1035 //------------------------------data_size---------------------------
1036 int SuperWord::data_size(Node* s) {
1037 int bsize = type2aelembytes(velt_basic_type(s));
1038 assert(bsize != 0, "valid size");
1039 return bsize;
1040 }
1041
1042 //------------------------------extend_packlist---------------------------
1043 // Extend packset by following use->def and def->use links from pack members.
1044 void SuperWord::extend_packlist() {
1045 bool changed;
1046 do {
1047 packset_sort(_packset.length());
1048 changed = false;
1049 for (int i = 0; i < _packset.length(); i++) {
1050 Node_List* p = _packset.at(i);
1051 changed |= follow_use_defs(p);
1052 changed |= follow_def_uses(p);
1053 }
1054 } while (changed);
1055
1056 if (_race_possible) {
1057 for (int i = 0; i < _packset.length(); i++) {
1058 Node_List* p = _packset.at(i);
1059 order_def_uses(p);
1060 }
1061 }
1062
1063 #ifndef PRODUCT
1064 if (TraceSuperWord) {
1065 tty->print_cr("\nAfter extend_packlist");
1066 print_packset();
1067 }
1068 #endif
1069 }
1070
1071 //------------------------------follow_use_defs---------------------------
1072 // Extend the packset by visiting operand definitions of nodes in pack p
1073 bool SuperWord::follow_use_defs(Node_List* p) {
1074 assert(p->size() == 2, "just checking");
1075 Node* s1 = p->at(0);
1076 Node* s2 = p->at(1);
1077 assert(s1->req() == s2->req(), "just checking");
1078 assert(alignment(s1) + data_size(s1) == alignment(s2), "just checking");
1079
1080 if (s1->is_Load()) return false;
1081
1082 int align = alignment(s1);
1083 bool changed = false;
1084 int start = s1->is_Store() ? MemNode::ValueIn : 1;
1085 int end = s1->is_Store() ? MemNode::ValueIn+1 : s1->req();
1086 for (int j = start; j < end; j++) {
1087 Node* t1 = s1->in(j);
1088 Node* t2 = s2->in(j);
1089 if (!in_bb(t1) || !in_bb(t2))
1090 continue;
1091 if (stmts_can_pack(t1, t2, align)) {
1092 if (est_savings(t1, t2) >= 0) {
1093 Node_List* pair = new Node_List();
1094 pair->push(t1);
1095 pair->push(t2);
1096 _packset.append(pair);
1097 set_alignment(t1, t2, align);
1098 changed = true;
1099 }
1100 }
1101 }
1102 return changed;
1103 }
1104
1105 //------------------------------follow_def_uses---------------------------
1106 // Extend the packset by visiting uses of nodes in pack p
1107 bool SuperWord::follow_def_uses(Node_List* p) {
1108 bool changed = false;
1109 Node* s1 = p->at(0);
1110 Node* s2 = p->at(1);
1111 assert(p->size() == 2, "just checking");
1112 assert(s1->req() == s2->req(), "just checking");
1113 assert(alignment(s1) + data_size(s1) == alignment(s2), "just checking");
1114
1115 if (s1->is_Store()) return false;
1116
1117 int align = alignment(s1);
1118 int savings = -1;
1119 int num_s1_uses = 0;
1120 Node* u1 = NULL;
1121 Node* u2 = NULL;
1122 for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
1123 Node* t1 = s1->fast_out(i);
1124 num_s1_uses++;
1125 if (!in_bb(t1)) continue;
1126 for (DUIterator_Fast jmax, j = s2->fast_outs(jmax); j < jmax; j++) {
1127 Node* t2 = s2->fast_out(j);
1128 if (!in_bb(t2)) continue;
1129 if (!opnd_positions_match(s1, t1, s2, t2))
1130 continue;
1131 if (stmts_can_pack(t1, t2, align)) {
1132 int my_savings = est_savings(t1, t2);
1133 if (my_savings > savings) {
1134 savings = my_savings;
1135 u1 = t1;
1136 u2 = t2;
1137 }
1138 }
1139 }
1140 }
1141 if (num_s1_uses > 1) {
1142 _race_possible = true;
1143 }
1144 if (savings >= 0) {
1145 Node_List* pair = new Node_List();
1146 pair->push(u1);
1147 pair->push(u2);
1148 _packset.append(pair);
1149 set_alignment(u1, u2, align);
1150 changed = true;
1151 }
1152 return changed;
1153 }
1154
1155 //------------------------------order_def_uses---------------------------
1156 // For extended packsets, ordinally arrange uses packset by major component
1157 void SuperWord::order_def_uses(Node_List* p) {
1158 Node* s1 = p->at(0);
1159
1160 if (s1->is_Store()) return;
1161
1162 // reductions are always managed beforehand
1163 if (s1->is_reduction()) return;
1164
1165 for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
1166 Node* t1 = s1->fast_out(i);
1167
1168 // Only allow operand swap on commuting operations
1169 if (!t1->is_Add() && !t1->is_Mul()) {
1170 break;
1171 }
1172
1173 // Now find t1's packset
1174 Node_List* p2 = NULL;
1175 for (int j = 0; j < _packset.length(); j++) {
1176 p2 = _packset.at(j);
1177 Node* first = p2->at(0);
1178 if (t1 == first) {
1179 break;
1180 }
1181 p2 = NULL;
1182 }
1183 // Arrange all sub components by the major component
1184 if (p2 != NULL) {
1185 for (uint j = 1; j < p->size(); j++) {
1186 Node* d1 = p->at(j);
1187 Node* u1 = p2->at(j);
1188 opnd_positions_match(s1, t1, d1, u1);
1189 }
1190 }
1191 }
1192 }
1193
1194 //---------------------------opnd_positions_match-------------------------
1195 // Is the use of d1 in u1 at the same operand position as d2 in u2?
1196 bool SuperWord::opnd_positions_match(Node* d1, Node* u1, Node* d2, Node* u2) {
1197 // check reductions to see if they are marshalled to represent the reduction
1198 // operator in a specified opnd
1199 if (u1->is_reduction() && u2->is_reduction()) {
1200 // ensure reductions have phis and reduction definitions feeding the 1st operand
1201 Node* first = u1->in(2);
1202 if (first->is_Phi() || first->is_reduction()) {
1203 u1->swap_edges(1, 2);
1204 }
1205 // ensure reductions have phis and reduction definitions feeding the 1st operand
1206 first = u2->in(2);
1207 if (first->is_Phi() || first->is_reduction()) {
1208 u2->swap_edges(1, 2);
1209 }
1210 return true;
1211 }
1212
1213 uint ct = u1->req();
1214 if (ct != u2->req()) return false;
1215 uint i1 = 0;
1216 uint i2 = 0;
1217 do {
1218 for (i1++; i1 < ct; i1++) if (u1->in(i1) == d1) break;
1219 for (i2++; i2 < ct; i2++) if (u2->in(i2) == d2) break;
1220 if (i1 != i2) {
1221 if ((i1 == (3-i2)) && (u2->is_Add() || u2->is_Mul())) {
1222 // Further analysis relies on operands position matching.
1223 u2->swap_edges(i1, i2);
1224 } else {
1225 return false;
1226 }
1227 }
1228 } while (i1 < ct);
1229 return true;
1230 }
1231
1232 //------------------------------est_savings---------------------------
1233 // Estimate the savings from executing s1 and s2 as a pack
1234 int SuperWord::est_savings(Node* s1, Node* s2) {
1235 int save_in = 2 - 1; // 2 operations per instruction in packed form
1236
1237 // inputs
1238 for (uint i = 1; i < s1->req(); i++) {
1239 Node* x1 = s1->in(i);
1240 Node* x2 = s2->in(i);
1241 if (x1 != x2) {
1242 if (are_adjacent_refs(x1, x2)) {
1243 save_in += adjacent_profit(x1, x2);
1244 } else if (!in_packset(x1, x2)) {
1245 save_in -= pack_cost(2);
1246 } else {
1247 save_in += unpack_cost(2);
1248 }
1249 }
1250 }
1251
1252 // uses of result
1253 uint ct = 0;
1254 int save_use = 0;
1255 for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
1256 Node* s1_use = s1->fast_out(i);
1257 for (int j = 0; j < _packset.length(); j++) {
1258 Node_List* p = _packset.at(j);
1259 if (p->at(0) == s1_use) {
1260 for (DUIterator_Fast kmax, k = s2->fast_outs(kmax); k < kmax; k++) {
1261 Node* s2_use = s2->fast_out(k);
1262 if (p->at(p->size()-1) == s2_use) {
1263 ct++;
1264 if (are_adjacent_refs(s1_use, s2_use)) {
1265 save_use += adjacent_profit(s1_use, s2_use);
1266 }
1267 }
1268 }
1269 }
1270 }
1271 }
1272
1273 if (ct < s1->outcnt()) save_use += unpack_cost(1);
1274 if (ct < s2->outcnt()) save_use += unpack_cost(1);
1275
1276 return MAX2(save_in, save_use);
1277 }
1278
1279 //------------------------------costs---------------------------
1280 int SuperWord::adjacent_profit(Node* s1, Node* s2) { return 2; }
1281 int SuperWord::pack_cost(int ct) { return ct; }
1282 int SuperWord::unpack_cost(int ct) { return ct; }
1283
1284 //------------------------------combine_packs---------------------------
1285 // Combine packs A and B with A.last == B.first into A.first..,A.last,B.second,..B.last
1286 void SuperWord::combine_packs() {
1287 bool changed = true;
1288 // Combine packs regardless max vector size.
1289 while (changed) {
1290 changed = false;
1291 for (int i = 0; i < _packset.length(); i++) {
1292 Node_List* p1 = _packset.at(i);
1293 if (p1 == NULL) continue;
1294 // Because of sorting we can start at i + 1
1295 for (int j = i + 1; j < _packset.length(); j++) {
1296 Node_List* p2 = _packset.at(j);
1297 if (p2 == NULL) continue;
1298 if (i == j) continue;
1299 if (p1->at(p1->size()-1) == p2->at(0)) {
1300 for (uint k = 1; k < p2->size(); k++) {
1301 p1->push(p2->at(k));
1302 }
1303 _packset.at_put(j, NULL);
1304 changed = true;
1305 }
1306 }
1307 }
1308 }
1309
1310 // Split packs which have size greater then max vector size.
1311 for (int i = 0; i < _packset.length(); i++) {
1312 Node_List* p1 = _packset.at(i);
1313 if (p1 != NULL) {
1314 BasicType bt = velt_basic_type(p1->at(0));
1315 uint max_vlen = Matcher::max_vector_size(bt); // Max elements in vector
1316 assert(is_power_of_2(max_vlen), "sanity");
1317 uint psize = p1->size();
1318 if (!is_power_of_2(psize)) {
1319 // Skip pack which can't be vector.
1320 // case1: for(...) { a[i] = i; } elements values are different (i+x)
1321 // case2: for(...) { a[i] = b[i+1]; } can't align both, load and store
1322 _packset.at_put(i, NULL);
1323 continue;
1324 }
1325 if (psize > max_vlen) {
1326 Node_List* pack = new Node_List();
1327 for (uint j = 0; j < psize; j++) {
1328 pack->push(p1->at(j));
1329 if (pack->size() >= max_vlen) {
1330 assert(is_power_of_2(pack->size()), "sanity");
1331 _packset.append(pack);
1332 pack = new Node_List();
1333 }
1334 }
1335 _packset.at_put(i, NULL);
1336 }
1337 }
1338 }
1339
1340 // Compress list.
1341 for (int i = _packset.length() - 1; i >= 0; i--) {
1342 Node_List* p1 = _packset.at(i);
1343 if (p1 == NULL) {
1344 _packset.remove_at(i);
1345 }
1346 }
1347
1348 #ifndef PRODUCT
1349 if (TraceSuperWord) {
1350 tty->print_cr("\nAfter combine_packs");
1351 print_packset();
1352 }
1353 #endif
1354 }
1355
1356 //-----------------------------construct_my_pack_map--------------------------
1357 // Construct the map from nodes to packs. Only valid after the
1358 // point where a node is only in one pack (after combine_packs).
1359 void SuperWord::construct_my_pack_map() {
1360 Node_List* rslt = NULL;
1361 for (int i = 0; i < _packset.length(); i++) {
1362 Node_List* p = _packset.at(i);
1363 for (uint j = 0; j < p->size(); j++) {
1364 Node* s = p->at(j);
1365 assert(my_pack(s) == NULL, "only in one pack");
1366 set_my_pack(s, p);
1367 }
1368 }
1369 }
1370
1371 //------------------------------filter_packs---------------------------
1372 // Remove packs that are not implemented or not profitable.
1373 void SuperWord::filter_packs() {
1374 // Remove packs that are not implemented
1375 for (int i = _packset.length() - 1; i >= 0; i--) {
1376 Node_List* pk = _packset.at(i);
1377 bool impl = implemented(pk);
1378 if (!impl) {
1379 #ifndef PRODUCT
1380 if (TraceSuperWord && Verbose) {
1381 tty->print_cr("Unimplemented");
1382 pk->at(0)->dump();
1383 }
1384 #endif
1385 remove_pack_at(i);
1386 }
1387 Node *n = pk->at(0);
1388 if (n->is_reduction()) {
1389 _num_reductions++;
1390 } else {
1391 _num_work_vecs++;
1392 }
1393 }
1394
1395 // Remove packs that are not profitable
1396 bool changed;
1397 do {
1398 changed = false;
1399 for (int i = _packset.length() - 1; i >= 0; i--) {
1400 Node_List* pk = _packset.at(i);
1401 bool prof = profitable(pk);
1402 if (!prof) {
1403 #ifndef PRODUCT
1404 if (TraceSuperWord && Verbose) {
1405 tty->print_cr("Unprofitable");
1406 pk->at(0)->dump();
1407 }
1408 #endif
1409 remove_pack_at(i);
1410 changed = true;
1411 }
1412 }
1413 } while (changed);
1414
1415 #ifndef PRODUCT
1416 if (TraceSuperWord) {
1417 tty->print_cr("\nAfter filter_packs");
1418 print_packset();
1419 tty->cr();
1420 }
1421 #endif
1422 }
1423
1424 //------------------------------implemented---------------------------
1425 // Can code be generated for pack p?
1426 bool SuperWord::implemented(Node_List* p) {
1427 bool retValue = false;
1428 Node* p0 = p->at(0);
1429 if (p0 != NULL) {
1430 int opc = p0->Opcode();
1431 uint size = p->size();
1432 if (p0->is_reduction()) {
1433 const Type *arith_type = p0->bottom_type();
1434 // Length 2 reductions of INT/LONG do not offer performance benefits
1435 if (((arith_type->basic_type() == T_INT) || (arith_type->basic_type() == T_LONG)) && (size == 2)) {
1436 retValue = false;
1437 } else {
1438 retValue = ReductionNode::implemented(opc, size, arith_type->basic_type());
1439 }
1440 } else {
1441 retValue = VectorNode::implemented(opc, size, velt_basic_type(p0));
1442 }
1443 }
1444 return retValue;
1445 }
1446
1447 //------------------------------same_inputs--------------------------
1448 // For pack p, are all idx operands the same?
1449 static bool same_inputs(Node_List* p, int idx) {
1450 Node* p0 = p->at(0);
1451 uint vlen = p->size();
1452 Node* p0_def = p0->in(idx);
1453 for (uint i = 1; i < vlen; i++) {
1454 Node* pi = p->at(i);
1455 Node* pi_def = pi->in(idx);
1456 if (p0_def != pi_def)
1457 return false;
1458 }
1459 return true;
1460 }
1461
1462 //------------------------------profitable---------------------------
1463 // For pack p, are all operands and all uses (with in the block) vector?
1464 bool SuperWord::profitable(Node_List* p) {
1465 Node* p0 = p->at(0);
1466 uint start, end;
1467 VectorNode::vector_operands(p0, &start, &end);
1468
1469 // Return false if some inputs are not vectors or vectors with different
1470 // size or alignment.
1471 // Also, for now, return false if not scalar promotion case when inputs are
1472 // the same. Later, implement PackNode and allow differing, non-vector inputs
1473 // (maybe just the ones from outside the block.)
1474 for (uint i = start; i < end; i++) {
1475 if (!is_vector_use(p0, i))
1476 return false;
1477 }
1478 // Check if reductions are connected
1479 if (p0->is_reduction()) {
1480 Node* second_in = p0->in(2);
1481 Node_List* second_pk = my_pack(second_in);
1482 if ((second_pk == NULL) || (_num_work_vecs == _num_reductions)) {
1483 // Remove reduction flag if no parent pack or if not enough work
1484 // to cover reduction expansion overhead
1485 p0->remove_flag(Node::Flag_is_reduction);
1486 return false;
1487 } else if (second_pk->size() != p->size()) {
1488 return false;
1489 }
1490 }
1491 if (VectorNode::is_shift(p0)) {
1492 // For now, return false if shift count is vector or not scalar promotion
1493 // case (different shift counts) because it is not supported yet.
1494 Node* cnt = p0->in(2);
1495 Node_List* cnt_pk = my_pack(cnt);
1496 if (cnt_pk != NULL)
1497 return false;
1498 if (!same_inputs(p, 2))
1499 return false;
1500 }
1501 if (!p0->is_Store()) {
1502 // For now, return false if not all uses are vector.
1503 // Later, implement ExtractNode and allow non-vector uses (maybe
1504 // just the ones outside the block.)
1505 for (uint i = 0; i < p->size(); i++) {
1506 Node* def = p->at(i);
1507 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
1508 Node* use = def->fast_out(j);
1509 for (uint k = 0; k < use->req(); k++) {
1510 Node* n = use->in(k);
1511 if (def == n) {
1512 // reductions can be loop carried dependences
1513 if (def->is_reduction() && use->is_Phi())
1514 continue;
1515 if (!is_vector_use(use, k)) {
1516 return false;
1517 }
1518 }
1519 }
1520 }
1521 }
1522 }
1523 return true;
1524 }
1525
1526 //------------------------------schedule---------------------------
1527 // Adjust the memory graph for the packed operations
1528 void SuperWord::schedule() {
1529
1530 // Co-locate in the memory graph the members of each memory pack
1531 for (int i = 0; i < _packset.length(); i++) {
1532 co_locate_pack(_packset.at(i));
1533 }
1534 }
1535
1536 //-------------------------------remove_and_insert-------------------
1537 // Remove "current" from its current position in the memory graph and insert
1538 // it after the appropriate insertion point (lip or uip).
1539 void SuperWord::remove_and_insert(MemNode *current, MemNode *prev, MemNode *lip,
1540 Node *uip, Unique_Node_List &sched_before) {
1541 Node* my_mem = current->in(MemNode::Memory);
1542 bool sched_up = sched_before.member(current);
1543
1544 // remove current_store from its current position in the memmory graph
1545 for (DUIterator i = current->outs(); current->has_out(i); i++) {
1546 Node* use = current->out(i);
1547 if (use->is_Mem()) {
1548 assert(use->in(MemNode::Memory) == current, "must be");
1549 if (use == prev) { // connect prev to my_mem
1550 _igvn.replace_input_of(use, MemNode::Memory, my_mem);
1551 --i; //deleted this edge; rescan position
1552 } else if (sched_before.member(use)) {
1553 if (!sched_up) { // Will be moved together with current
1554 _igvn.replace_input_of(use, MemNode::Memory, uip);
1555 --i; //deleted this edge; rescan position
1556 }
1557 } else {
1558 if (sched_up) { // Will be moved together with current
1559 _igvn.replace_input_of(use, MemNode::Memory, lip);
1560 --i; //deleted this edge; rescan position
1561 }
1562 }
1563 }
1564 }
1565
1566 Node *insert_pt = sched_up ? uip : lip;
1567
1568 // all uses of insert_pt's memory state should use current's instead
1569 for (DUIterator i = insert_pt->outs(); insert_pt->has_out(i); i++) {
1570 Node* use = insert_pt->out(i);
1571 if (use->is_Mem()) {
1572 assert(use->in(MemNode::Memory) == insert_pt, "must be");
1573 _igvn.replace_input_of(use, MemNode::Memory, current);
1574 --i; //deleted this edge; rescan position
1575 } else if (!sched_up && use->is_Phi() && use->bottom_type() == Type::MEMORY) {
1576 uint pos; //lip (lower insert point) must be the last one in the memory slice
1577 for (pos=1; pos < use->req(); pos++) {
1578 if (use->in(pos) == insert_pt) break;
1579 }
1580 _igvn.replace_input_of(use, pos, current);
1581 --i;
1582 }
1583 }
1584
1585 //connect current to insert_pt
1586 _igvn.replace_input_of(current, MemNode::Memory, insert_pt);
1587 }
1588
1589 //------------------------------co_locate_pack----------------------------------
1590 // To schedule a store pack, we need to move any sandwiched memory ops either before
1591 // or after the pack, based upon dependence information:
1592 // (1) If any store in the pack depends on the sandwiched memory op, the
1593 // sandwiched memory op must be scheduled BEFORE the pack;
1594 // (2) If a sandwiched memory op depends on any store in the pack, the
1595 // sandwiched memory op must be scheduled AFTER the pack;
1596 // (3) If a sandwiched memory op (say, memA) depends on another sandwiched
1597 // memory op (say memB), memB must be scheduled before memA. So, if memA is
1598 // scheduled before the pack, memB must also be scheduled before the pack;
1599 // (4) If there is no dependence restriction for a sandwiched memory op, we simply
1600 // schedule this store AFTER the pack
1601 // (5) We know there is no dependence cycle, so there in no other case;
1602 // (6) Finally, all memory ops in another single pack should be moved in the same direction.
1603 //
1604 // To schedule a load pack, we use the memory state of either the first or the last load in
1605 // the pack, based on the dependence constraint.
1606 void SuperWord::co_locate_pack(Node_List* pk) {
1607 if (pk->at(0)->is_Store()) {
1608 MemNode* first = executed_first(pk)->as_Mem();
1609 MemNode* last = executed_last(pk)->as_Mem();
1610 Unique_Node_List schedule_before_pack;
1611 Unique_Node_List memops;
1612
1613 MemNode* current = last->in(MemNode::Memory)->as_Mem();
1614 MemNode* previous = last;
1615 while (true) {
1616 assert(in_bb(current), "stay in block");
1617 memops.push(previous);
1618 for (DUIterator i = current->outs(); current->has_out(i); i++) {
1619 Node* use = current->out(i);
1620 if (use->is_Mem() && use != previous)
1621 memops.push(use);
1622 }
1623 if (current == first) break;
1624 previous = current;
1625 current = current->in(MemNode::Memory)->as_Mem();
1626 }
1627
1628 // determine which memory operations should be scheduled before the pack
1629 for (uint i = 1; i < memops.size(); i++) {
1630 Node *s1 = memops.at(i);
1631 if (!in_pack(s1, pk) && !schedule_before_pack.member(s1)) {
1632 for (uint j = 0; j< i; j++) {
1633 Node *s2 = memops.at(j);
1634 if (!independent(s1, s2)) {
1635 if (in_pack(s2, pk) || schedule_before_pack.member(s2)) {
1636 schedule_before_pack.push(s1); // s1 must be scheduled before
1637 Node_List* mem_pk = my_pack(s1);
1638 if (mem_pk != NULL) {
1639 for (uint ii = 0; ii < mem_pk->size(); ii++) {
1640 Node* s = mem_pk->at(ii); // follow partner
1641 if (memops.member(s) && !schedule_before_pack.member(s))
1642 schedule_before_pack.push(s);
1643 }
1644 }
1645 break;
1646 }
1647 }
1648 }
1649 }
1650 }
1651
1652 Node* upper_insert_pt = first->in(MemNode::Memory);
1653 // Following code moves loads connected to upper_insert_pt below aliased stores.
1654 // Collect such loads here and reconnect them back to upper_insert_pt later.
1655 memops.clear();
1656 for (DUIterator i = upper_insert_pt->outs(); upper_insert_pt->has_out(i); i++) {
1657 Node* use = upper_insert_pt->out(i);
1658 if (use->is_Mem() && !use->is_Store()) {
1659 memops.push(use);
1660 }
1661 }
1662
1663 MemNode* lower_insert_pt = last;
1664 previous = last; //previous store in pk
1665 current = last->in(MemNode::Memory)->as_Mem();
1666
1667 // start scheduling from "last" to "first"
1668 while (true) {
1669 assert(in_bb(current), "stay in block");
1670 assert(in_pack(previous, pk), "previous stays in pack");
1671 Node* my_mem = current->in(MemNode::Memory);
1672
1673 if (in_pack(current, pk)) {
1674 // Forward users of my memory state (except "previous) to my input memory state
1675 for (DUIterator i = current->outs(); current->has_out(i); i++) {
1676 Node* use = current->out(i);
1677 if (use->is_Mem() && use != previous) {
1678 assert(use->in(MemNode::Memory) == current, "must be");
1679 if (schedule_before_pack.member(use)) {
1680 _igvn.replace_input_of(use, MemNode::Memory, upper_insert_pt);
1681 } else {
1682 _igvn.replace_input_of(use, MemNode::Memory, lower_insert_pt);
1683 }
1684 --i; // deleted this edge; rescan position
1685 }
1686 }
1687 previous = current;
1688 } else { // !in_pack(current, pk) ==> a sandwiched store
1689 remove_and_insert(current, previous, lower_insert_pt, upper_insert_pt, schedule_before_pack);
1690 }
1691
1692 if (current == first) break;
1693 current = my_mem->as_Mem();
1694 } // end while
1695
1696 // Reconnect loads back to upper_insert_pt.
1697 for (uint i = 0; i < memops.size(); i++) {
1698 Node *ld = memops.at(i);
1699 if (ld->in(MemNode::Memory) != upper_insert_pt) {
1700 _igvn.replace_input_of(ld, MemNode::Memory, upper_insert_pt);
1701 }
1702 }
1703 } else if (pk->at(0)->is_Load()) { //load
1704 // all loads in the pack should have the same memory state. By default,
1705 // we use the memory state of the last load. However, if any load could
1706 // not be moved down due to the dependence constraint, we use the memory
1707 // state of the first load.
1708 Node* last_mem = executed_last(pk)->in(MemNode::Memory);
1709 Node* first_mem = executed_first(pk)->in(MemNode::Memory);
1710 bool schedule_last = true;
1711 for (uint i = 0; i < pk->size(); i++) {
1712 Node* ld = pk->at(i);
1713 for (Node* current = last_mem; current != ld->in(MemNode::Memory);
1714 current=current->in(MemNode::Memory)) {
1715 assert(current != first_mem, "corrupted memory graph");
1716 if(current->is_Mem() && !independent(current, ld)){
1717 schedule_last = false; // a later store depends on this load
1718 break;
1719 }
1720 }
1721 }
1722
1723 Node* mem_input = schedule_last ? last_mem : first_mem;
1724 _igvn.hash_delete(mem_input);
1725 // Give each load the same memory state
1726 for (uint i = 0; i < pk->size(); i++) {
1727 LoadNode* ld = pk->at(i)->as_Load();
1728 _igvn.replace_input_of(ld, MemNode::Memory, mem_input);
1729 }
1730 }
1731 }
1732
1733 //------------------------------output---------------------------
1734 // Convert packs into vector node operations
1735 void SuperWord::output() {
1736 if (_packset.length() == 0) return;
1737
1738 #ifndef PRODUCT
1739 if (TraceLoopOpts) {
1740 tty->print("SuperWord ");
1741 lpt()->dump_head();
1742 }
1743 #endif
1744
1745 // MUST ENSURE main loop's initial value is properly aligned:
1746 // (iv_initial_value + min_iv_offset) % vector_width_in_bytes() == 0
1747
1748 align_initial_loop_index(align_to_ref());
1749
1750 // Insert extract (unpack) operations for scalar uses
1751 for (int i = 0; i < _packset.length(); i++) {
1752 insert_extracts(_packset.at(i));
1753 }
1754
1755 Compile* C = _phase->C;
1756 uint max_vlen_in_bytes = 0;
1757 for (int i = 0; i < _block.length(); i++) {
1758 Node* n = _block.at(i);
1759 Node_List* p = my_pack(n);
1760 if (p && n == executed_last(p)) {
1761 uint vlen = p->size();
1762 uint vlen_in_bytes = 0;
1763 Node* vn = NULL;
1764 Node* low_adr = p->at(0);
1765 Node* first = executed_first(p);
1766 int opc = n->Opcode();
1767 if (n->is_Load()) {
1768 Node* ctl = n->in(MemNode::Control);
1769 Node* mem = first->in(MemNode::Memory);
1770 SWPointer p1(n->as_Mem(), this, NULL, false);
1771 // Identify the memory dependency for the new loadVector node by
1772 // walking up through memory chain.
1773 // This is done to give flexibility to the new loadVector node so that
1774 // it can move above independent storeVector nodes.
1775 while (mem->is_StoreVector()) {
1776 SWPointer p2(mem->as_Mem(), this, NULL, false);
1777 int cmp = p1.cmp(p2);
1778 if (SWPointer::not_equal(cmp) || !SWPointer::comparable(cmp)) {
1779 mem = mem->in(MemNode::Memory);
1780 } else {
1781 break; // dependent memory
1782 }
1783 }
1784 Node* adr = low_adr->in(MemNode::Address);
1785 const TypePtr* atyp = n->adr_type();
1786 vn = LoadVectorNode::make(opc, ctl, mem, adr, atyp, vlen, velt_basic_type(n));
1787 vlen_in_bytes = vn->as_LoadVector()->memory_size();
1788 } else if (n->is_Store()) {
1789 // Promote value to be stored to vector
1790 Node* val = vector_opd(p, MemNode::ValueIn);
1791 Node* ctl = n->in(MemNode::Control);
1792 Node* mem = first->in(MemNode::Memory);
1793 Node* adr = low_adr->in(MemNode::Address);
1794 const TypePtr* atyp = n->adr_type();
1795 vn = StoreVectorNode::make(opc, ctl, mem, adr, atyp, val, vlen);
1796 vlen_in_bytes = vn->as_StoreVector()->memory_size();
1797 } else if (n->req() == 3) {
1798 // Promote operands to vector
1799 Node* in1 = NULL;
1800 bool node_isa_reduction = n->is_reduction();
1801 if (node_isa_reduction) {
1802 // the input to the first reduction operation is retained
1803 in1 = low_adr->in(1);
1804 } else {
1805 in1 = vector_opd(p, 1);
1806 }
1807 Node* in2 = vector_opd(p, 2);
1808 if (VectorNode::is_invariant_vector(in1) && (node_isa_reduction == false) && (n->is_Add() || n->is_Mul())) {
1809 // Move invariant vector input into second position to avoid register spilling.
1810 Node* tmp = in1;
1811 in1 = in2;
1812 in2 = tmp;
1813 }
1814 if (node_isa_reduction) {
1815 const Type *arith_type = n->bottom_type();
1816 vn = ReductionNode::make(opc, NULL, in1, in2, arith_type->basic_type());
1817 if (in2->is_Load()) {
1818 vlen_in_bytes = in2->as_LoadVector()->memory_size();
1819 } else {
1820 vlen_in_bytes = in2->as_Vector()->length_in_bytes();
1821 }
1822 } else {
1823 vn = VectorNode::make(opc, in1, in2, vlen, velt_basic_type(n));
1824 vlen_in_bytes = vn->as_Vector()->length_in_bytes();
1825 }
1826 } else {
1827 ShouldNotReachHere();
1828 }
1829 assert(vn != NULL, "sanity");
1830 _igvn.register_new_node_with_optimizer(vn);
1831 _phase->set_ctrl(vn, _phase->get_ctrl(p->at(0)));
1832 for (uint j = 0; j < p->size(); j++) {
1833 Node* pm = p->at(j);
1834 _igvn.replace_node(pm, vn);
1835 }
1836 _igvn._worklist.push(vn);
1837
1838 if (vlen_in_bytes > max_vlen_in_bytes) {
1839 max_vlen_in_bytes = vlen_in_bytes;
1840 }
1841 #ifdef ASSERT
1842 if (TraceNewVectors) {
1843 tty->print("new Vector node: ");
1844 vn->dump();
1845 }
1846 #endif
1847 }
1848 }
1849 C->set_max_vector_size(max_vlen_in_bytes);
1850 }
1851
1852 //------------------------------vector_opd---------------------------
1853 // Create a vector operand for the nodes in pack p for operand: in(opd_idx)
1854 Node* SuperWord::vector_opd(Node_List* p, int opd_idx) {
1855 Node* p0 = p->at(0);
1856 uint vlen = p->size();
1857 Node* opd = p0->in(opd_idx);
1858
1859 if (same_inputs(p, opd_idx)) {
1860 if (opd->is_Vector() || opd->is_LoadVector()) {
1861 assert(((opd_idx != 2) || !VectorNode::is_shift(p0)), "shift's count can't be vector");
1862 return opd; // input is matching vector
1863 }
1864 if ((opd_idx == 2) && VectorNode::is_shift(p0)) {
1865 Compile* C = _phase->C;
1866 Node* cnt = opd;
1867 // Vector instructions do not mask shift count, do it here.
1868 juint mask = (p0->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
1869 const TypeInt* t = opd->find_int_type();
1870 if (t != NULL && t->is_con()) {
1871 juint shift = t->get_con();
1872 if (shift > mask) { // Unsigned cmp
1873 cnt = ConNode::make(TypeInt::make(shift & mask));
1874 }
1875 } else {
1876 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
1877 cnt = ConNode::make(TypeInt::make(mask));
1878 _igvn.register_new_node_with_optimizer(cnt);
1879 cnt = new AndINode(opd, cnt);
1880 _igvn.register_new_node_with_optimizer(cnt);
1881 _phase->set_ctrl(cnt, _phase->get_ctrl(opd));
1882 }
1883 assert(opd->bottom_type()->isa_int(), "int type only");
1884 // Move non constant shift count into vector register.
1885 cnt = VectorNode::shift_count(p0, cnt, vlen, velt_basic_type(p0));
1886 }
1887 if (cnt != opd) {
1888 _igvn.register_new_node_with_optimizer(cnt);
1889 _phase->set_ctrl(cnt, _phase->get_ctrl(opd));
1890 }
1891 return cnt;
1892 }
1893 assert(!opd->is_StoreVector(), "such vector is not expected here");
1894 // Convert scalar input to vector with the same number of elements as
1895 // p0's vector. Use p0's type because size of operand's container in
1896 // vector should match p0's size regardless operand's size.
1897 const Type* p0_t = velt_type(p0);
1898 VectorNode* vn = VectorNode::scalar2vector(opd, vlen, p0_t);
1899
1900 _igvn.register_new_node_with_optimizer(vn);
1901 _phase->set_ctrl(vn, _phase->get_ctrl(opd));
1902 #ifdef ASSERT
1903 if (TraceNewVectors) {
1904 tty->print("new Vector node: ");
1905 vn->dump();
1906 }
1907 #endif
1908 return vn;
1909 }
1910
1911 // Insert pack operation
1912 BasicType bt = velt_basic_type(p0);
1913 PackNode* pk = PackNode::make(opd, vlen, bt);
1914 DEBUG_ONLY( const BasicType opd_bt = opd->bottom_type()->basic_type(); )
1915
1916 for (uint i = 1; i < vlen; i++) {
1917 Node* pi = p->at(i);
1918 Node* in = pi->in(opd_idx);
1919 assert(my_pack(in) == NULL, "Should already have been unpacked");
1920 assert(opd_bt == in->bottom_type()->basic_type(), "all same type");
1921 pk->add_opd(in);
1922 }
1923 _igvn.register_new_node_with_optimizer(pk);
1924 _phase->set_ctrl(pk, _phase->get_ctrl(opd));
1925 #ifdef ASSERT
1926 if (TraceNewVectors) {
1927 tty->print("new Vector node: ");
1928 pk->dump();
1929 }
1930 #endif
1931 return pk;
1932 }
1933
1934 //------------------------------insert_extracts---------------------------
1935 // If a use of pack p is not a vector use, then replace the
1936 // use with an extract operation.
1937 void SuperWord::insert_extracts(Node_List* p) {
1938 if (p->at(0)->is_Store()) return;
1939 assert(_n_idx_list.is_empty(), "empty (node,index) list");
1940
1941 // Inspect each use of each pack member. For each use that is
1942 // not a vector use, replace the use with an extract operation.
1943
1944 for (uint i = 0; i < p->size(); i++) {
1945 Node* def = p->at(i);
1946 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
1947 Node* use = def->fast_out(j);
1948 for (uint k = 0; k < use->req(); k++) {
1949 Node* n = use->in(k);
1950 if (def == n) {
1951 if (!is_vector_use(use, k)) {
1952 _n_idx_list.push(use, k);
1953 }
1954 }
1955 }
1956 }
1957 }
1958
1959 while (_n_idx_list.is_nonempty()) {
1960 Node* use = _n_idx_list.node();
1961 int idx = _n_idx_list.index();
1962 _n_idx_list.pop();
1963 Node* def = use->in(idx);
1964
1965 if (def->is_reduction()) continue;
1966
1967 // Insert extract operation
1968 _igvn.hash_delete(def);
1969 int def_pos = alignment(def) / data_size(def);
1970
1971 Node* ex = ExtractNode::make(def, def_pos, velt_basic_type(def));
1972 _igvn.register_new_node_with_optimizer(ex);
1973 _phase->set_ctrl(ex, _phase->get_ctrl(def));
1974 _igvn.replace_input_of(use, idx, ex);
1975 _igvn._worklist.push(def);
1976
1977 bb_insert_after(ex, bb_idx(def));
1978 set_velt_type(ex, velt_type(def));
1979 }
1980 }
1981
1982 //------------------------------is_vector_use---------------------------
1983 // Is use->in(u_idx) a vector use?
1984 bool SuperWord::is_vector_use(Node* use, int u_idx) {
1985 Node_List* u_pk = my_pack(use);
1986 if (u_pk == NULL) return false;
1987 if (use->is_reduction()) return true;
1988 Node* def = use->in(u_idx);
1989 Node_List* d_pk = my_pack(def);
1990 if (d_pk == NULL) {
1991 // check for scalar promotion
1992 Node* n = u_pk->at(0)->in(u_idx);
1993 for (uint i = 1; i < u_pk->size(); i++) {
1994 if (u_pk->at(i)->in(u_idx) != n) return false;
1995 }
1996 return true;
1997 }
1998 if (u_pk->size() != d_pk->size())
1999 return false;
2000 for (uint i = 0; i < u_pk->size(); i++) {
2001 Node* ui = u_pk->at(i);
2002 Node* di = d_pk->at(i);
2003 if (ui->in(u_idx) != di || alignment(ui) != alignment(di))
2004 return false;
2005 }
2006 return true;
2007 }
2008
2009 //------------------------------construct_bb---------------------------
2010 // Construct reverse postorder list of block members
2011 bool SuperWord::construct_bb() {
2012 Node* entry = bb();
2013
2014 assert(_stk.length() == 0, "stk is empty");
2015 assert(_block.length() == 0, "block is empty");
2016 assert(_data_entry.length() == 0, "data_entry is empty");
2017 assert(_mem_slice_head.length() == 0, "mem_slice_head is empty");
2018 assert(_mem_slice_tail.length() == 0, "mem_slice_tail is empty");
2019
2020 // Find non-control nodes with no inputs from within block,
2021 // create a temporary map from node _idx to bb_idx for use
2022 // by the visited and post_visited sets,
2023 // and count number of nodes in block.
2024 int bb_ct = 0;
2025 for (uint i = 0; i < lpt()->_body.size(); i++) {
2026 Node *n = lpt()->_body.at(i);
2027 set_bb_idx(n, i); // Create a temporary map
2028 if (in_bb(n)) {
2029 if (n->is_LoadStore() || n->is_MergeMem() ||
2030 (n->is_Proj() && !n->as_Proj()->is_CFG())) {
2031 // Bailout if the loop has LoadStore, MergeMem or data Proj
2032 // nodes. Superword optimization does not work with them.
2033 return false;
2034 }
2035 bb_ct++;
2036 if (!n->is_CFG()) {
2037 bool found = false;
2038 for (uint j = 0; j < n->req(); j++) {
2039 Node* def = n->in(j);
2040 if (def && in_bb(def)) {
2041 found = true;
2042 break;
2043 }
2044 }
2045 if (!found) {
2046 assert(n != entry, "can't be entry");
2047 _data_entry.push(n);
2048 }
2049 }
2050 }
2051 }
2052
2053 // Find memory slices (head and tail)
2054 for (DUIterator_Fast imax, i = lp()->fast_outs(imax); i < imax; i++) {
2055 Node *n = lp()->fast_out(i);
2056 if (in_bb(n) && (n->is_Phi() && n->bottom_type() == Type::MEMORY)) {
2057 Node* n_tail = n->in(LoopNode::LoopBackControl);
2058 if (n_tail != n->in(LoopNode::EntryControl)) {
2059 if (!n_tail->is_Mem()) {
2060 assert(n_tail->is_Mem(), err_msg_res("unexpected node for memory slice: %s", n_tail->Name()));
2061 return false; // Bailout
2062 }
2063 _mem_slice_head.push(n);
2064 _mem_slice_tail.push(n_tail);
2065 }
2066 }
2067 }
2068
2069 // Create an RPO list of nodes in block
2070
2071 visited_clear();
2072 post_visited_clear();
2073
2074 // Push all non-control nodes with no inputs from within block, then control entry
2075 for (int j = 0; j < _data_entry.length(); j++) {
2076 Node* n = _data_entry.at(j);
2077 visited_set(n);
2078 _stk.push(n);
2079 }
2080 visited_set(entry);
2081 _stk.push(entry);
2082
2083 // Do a depth first walk over out edges
2084 int rpo_idx = bb_ct - 1;
2085 int size;
2086 int reduction_uses = 0;
2087 while ((size = _stk.length()) > 0) {
2088 Node* n = _stk.top(); // Leave node on stack
2089 if (!visited_test_set(n)) {
2090 // forward arc in graph
2091 } else if (!post_visited_test(n)) {
2092 // cross or back arc
2093 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2094 Node *use = n->fast_out(i);
2095 if (in_bb(use) && !visited_test(use) &&
2096 // Don't go around backedge
2097 (!use->is_Phi() || n == entry)) {
2098 if (use->is_reduction()) {
2099 // First see if we can map the reduction on the given system we are on, then
2100 // make a data entry operation for each reduction we see.
2101 BasicType bt = use->bottom_type()->basic_type();
2102 if (ReductionNode::implemented(use->Opcode(), Matcher::min_vector_size(bt), bt)) {
2103 reduction_uses++;
2104 }
2105 }
2106 _stk.push(use);
2107 }
2108 }
2109 if (_stk.length() == size) {
2110 // There were no additional uses, post visit node now
2111 _stk.pop(); // Remove node from stack
2112 assert(rpo_idx >= 0, "");
2113 _block.at_put_grow(rpo_idx, n);
2114 rpo_idx--;
2115 post_visited_set(n);
2116 assert(rpo_idx >= 0 || _stk.is_empty(), "");
2117 }
2118 } else {
2119 _stk.pop(); // Remove post-visited node from stack
2120 }
2121 }
2122
2123 // Create real map of block indices for nodes
2124 for (int j = 0; j < _block.length(); j++) {
2125 Node* n = _block.at(j);
2126 set_bb_idx(n, j);
2127 }
2128
2129 // Ensure extra info is allocated.
2130 initialize_bb();
2131
2132 #ifndef PRODUCT
2133 if (TraceSuperWord) {
2134 print_bb();
2135 tty->print_cr("\ndata entry nodes: %s", _data_entry.length() > 0 ? "" : "NONE");
2136 for (int m = 0; m < _data_entry.length(); m++) {
2137 tty->print("%3d ", m);
2138 _data_entry.at(m)->dump();
2139 }
2140 tty->print_cr("\nmemory slices: %s", _mem_slice_head.length() > 0 ? "" : "NONE");
2141 for (int m = 0; m < _mem_slice_head.length(); m++) {
2142 tty->print("%3d ", m); _mem_slice_head.at(m)->dump();
2143 tty->print(" "); _mem_slice_tail.at(m)->dump();
2144 }
2145 }
2146 #endif
2147 assert(rpo_idx == -1 && bb_ct == _block.length(), "all block members found");
2148 return (_mem_slice_head.length() > 0) || (reduction_uses > 0) || (_data_entry.length() > 0);
2149 }
2150
2151 //------------------------------initialize_bb---------------------------
2152 // Initialize per node info
2153 void SuperWord::initialize_bb() {
2154 Node* last = _block.at(_block.length() - 1);
2155 grow_node_info(bb_idx(last));
2156 }
2157
2158 //------------------------------bb_insert_after---------------------------
2159 // Insert n into block after pos
2160 void SuperWord::bb_insert_after(Node* n, int pos) {
2161 int n_pos = pos + 1;
2162 // Make room
2163 for (int i = _block.length() - 1; i >= n_pos; i--) {
2164 _block.at_put_grow(i+1, _block.at(i));
2165 }
2166 for (int j = _node_info.length() - 1; j >= n_pos; j--) {
2167 _node_info.at_put_grow(j+1, _node_info.at(j));
2168 }
2169 // Set value
2170 _block.at_put_grow(n_pos, n);
2171 _node_info.at_put_grow(n_pos, SWNodeInfo::initial);
2172 // Adjust map from node->_idx to _block index
2173 for (int i = n_pos; i < _block.length(); i++) {
2174 set_bb_idx(_block.at(i), i);
2175 }
2176 }
2177
2178 //------------------------------compute_max_depth---------------------------
2179 // Compute max depth for expressions from beginning of block
2180 // Use to prune search paths during test for independence.
2181 void SuperWord::compute_max_depth() {
2182 int ct = 0;
2183 bool again;
2184 do {
2185 again = false;
2186 for (int i = 0; i < _block.length(); i++) {
2187 Node* n = _block.at(i);
2188 if (!n->is_Phi()) {
2189 int d_orig = depth(n);
2190 int d_in = 0;
2191 for (DepPreds preds(n, _dg); !preds.done(); preds.next()) {
2192 Node* pred = preds.current();
2193 if (in_bb(pred)) {
2194 d_in = MAX2(d_in, depth(pred));
2195 }
2196 }
2197 if (d_in + 1 != d_orig) {
2198 set_depth(n, d_in + 1);
2199 again = true;
2200 }
2201 }
2202 }
2203 ct++;
2204 } while (again);
2205 #ifndef PRODUCT
2206 if (TraceSuperWord && Verbose)
2207 tty->print_cr("compute_max_depth iterated: %d times", ct);
2208 #endif
2209 }
2210
2211 //-------------------------compute_vector_element_type-----------------------
2212 // Compute necessary vector element type for expressions
2213 // This propagates backwards a narrower integer type when the
2214 // upper bits of the value are not needed.
2215 // Example: char a,b,c; a = b + c;
2216 // Normally the type of the add is integer, but for packed character
2217 // operations the type of the add needs to be char.
2218 void SuperWord::compute_vector_element_type() {
2219 #ifndef PRODUCT
2220 if (TraceSuperWord && Verbose)
2221 tty->print_cr("\ncompute_velt_type:");
2222 #endif
2223
2224 // Initial type
2225 for (int i = 0; i < _block.length(); i++) {
2226 Node* n = _block.at(i);
2227 set_velt_type(n, container_type(n));
2228 }
2229
2230 // Propagate integer narrowed type backwards through operations
2231 // that don't depend on higher order bits
2232 for (int i = _block.length() - 1; i >= 0; i--) {
2233 Node* n = _block.at(i);
2234 // Only integer types need be examined
2235 const Type* vtn = velt_type(n);
2236 if (vtn->basic_type() == T_INT) {
2237 uint start, end;
2238 VectorNode::vector_operands(n, &start, &end);
2239
2240 for (uint j = start; j < end; j++) {
2241 Node* in = n->in(j);
2242 // Don't propagate through a memory
2243 if (!in->is_Mem() && in_bb(in) && velt_type(in)->basic_type() == T_INT &&
2244 data_size(n) < data_size(in)) {
2245 bool same_type = true;
2246 for (DUIterator_Fast kmax, k = in->fast_outs(kmax); k < kmax; k++) {
2247 Node *use = in->fast_out(k);
2248 if (!in_bb(use) || !same_velt_type(use, n)) {
2249 same_type = false;
2250 break;
2251 }
2252 }
2253 if (same_type) {
2254 // For right shifts of small integer types (bool, byte, char, short)
2255 // we need precise information about sign-ness. Only Load nodes have
2256 // this information because Store nodes are the same for signed and
2257 // unsigned values. And any arithmetic operation after a load may
2258 // expand a value to signed Int so such right shifts can't be used
2259 // because vector elements do not have upper bits of Int.
2260 const Type* vt = vtn;
2261 if (VectorNode::is_shift(in)) {
2262 Node* load = in->in(1);
2263 if (load->is_Load() && in_bb(load) && (velt_type(load)->basic_type() == T_INT)) {
2264 vt = velt_type(load);
2265 } else if (in->Opcode() != Op_LShiftI) {
2266 // Widen type to Int to avoid creation of right shift vector
2267 // (align + data_size(s1) check in stmts_can_pack() will fail).
2268 // Note, left shifts work regardless type.
2269 vt = TypeInt::INT;
2270 }
2271 }
2272 set_velt_type(in, vt);
2273 }
2274 }
2275 }
2276 }
2277 }
2278 #ifndef PRODUCT
2279 if (TraceSuperWord && Verbose) {
2280 for (int i = 0; i < _block.length(); i++) {
2281 Node* n = _block.at(i);
2282 velt_type(n)->dump();
2283 tty->print("\t");
2284 n->dump();
2285 }
2286 }
2287 #endif
2288 }
2289
2290 //------------------------------memory_alignment---------------------------
2291 // Alignment within a vector memory reference
2292 int SuperWord::memory_alignment(MemNode* s, int iv_adjust) {
2293 SWPointer p(s, this, NULL, false);
2294 if (!p.valid()) {
2295 return bottom_align;
2296 }
2297 int vw = vector_width_in_bytes(s);
2298 if (vw < 2) {
2299 return bottom_align; // No vectors for this type
2300 }
2301 int offset = p.offset_in_bytes();
2302 offset += iv_adjust*p.memory_size();
2303 int off_rem = offset % vw;
2304 int off_mod = off_rem >= 0 ? off_rem : off_rem + vw;
2305 return off_mod;
2306 }
2307
2308 //---------------------------container_type---------------------------
2309 // Smallest type containing range of values
2310 const Type* SuperWord::container_type(Node* n) {
2311 if (n->is_Mem()) {
2312 BasicType bt = n->as_Mem()->memory_type();
2313 if (n->is_Store() && (bt == T_CHAR)) {
2314 // Use T_SHORT type instead of T_CHAR for stored values because any
2315 // preceding arithmetic operation extends values to signed Int.
2316 bt = T_SHORT;
2317 }
2318 if (n->Opcode() == Op_LoadUB) {
2319 // Adjust type for unsigned byte loads, it is important for right shifts.
2320 // T_BOOLEAN is used because there is no basic type representing type
2321 // TypeInt::UBYTE. Use of T_BOOLEAN for vectors is fine because only
2322 // size (one byte) and sign is important.
2323 bt = T_BOOLEAN;
2324 }
2325 return Type::get_const_basic_type(bt);
2326 }
2327 const Type* t = _igvn.type(n);
2328 if (t->basic_type() == T_INT) {
2329 // A narrow type of arithmetic operations will be determined by
2330 // propagating the type of memory operations.
2331 return TypeInt::INT;
2332 }
2333 return t;
2334 }
2335
2336 bool SuperWord::same_velt_type(Node* n1, Node* n2) {
2337 const Type* vt1 = velt_type(n1);
2338 const Type* vt2 = velt_type(n2);
2339 if (vt1->basic_type() == T_INT && vt2->basic_type() == T_INT) {
2340 // Compare vectors element sizes for integer types.
2341 return data_size(n1) == data_size(n2);
2342 }
2343 return vt1 == vt2;
2344 }
2345
2346 //------------------------------in_packset---------------------------
2347 // Are s1 and s2 in a pack pair and ordered as s1,s2?
2348 bool SuperWord::in_packset(Node* s1, Node* s2) {
2349 for (int i = 0; i < _packset.length(); i++) {
2350 Node_List* p = _packset.at(i);
2351 assert(p->size() == 2, "must be");
2352 if (p->at(0) == s1 && p->at(p->size()-1) == s2) {
2353 return true;
2354 }
2355 }
2356 return false;
2357 }
2358
2359 //------------------------------in_pack---------------------------
2360 // Is s in pack p?
2361 Node_List* SuperWord::in_pack(Node* s, Node_List* p) {
2362 for (uint i = 0; i < p->size(); i++) {
2363 if (p->at(i) == s) {
2364 return p;
2365 }
2366 }
2367 return NULL;
2368 }
2369
2370 //------------------------------remove_pack_at---------------------------
2371 // Remove the pack at position pos in the packset
2372 void SuperWord::remove_pack_at(int pos) {
2373 Node_List* p = _packset.at(pos);
2374 for (uint i = 0; i < p->size(); i++) {
2375 Node* s = p->at(i);
2376 set_my_pack(s, NULL);
2377 }
2378 _packset.remove_at(pos);
2379 }
2380
2381 void SuperWord::packset_sort(int n) {
2382 // simple bubble sort so that we capitalize with O(n) when its already sorted
2383 while (n != 0) {
2384 bool swapped = false;
2385 for (int i = 1; i < n; i++) {
2386 Node_List* q_low = _packset.at(i-1);
2387 Node_List* q_i = _packset.at(i);
2388
2389 // only swap when we find something to swap
2390 if (alignment(q_low->at(0)) > alignment(q_i->at(0))) {
2391 Node_List* t = q_i;
2392 *(_packset.adr_at(i)) = q_low;
2393 *(_packset.adr_at(i-1)) = q_i;
2394 swapped = true;
2395 }
2396 }
2397 if (swapped == false) break;
2398 n--;
2399 }
2400 }
2401
2402 //------------------------------executed_first---------------------------
2403 // Return the node executed first in pack p. Uses the RPO block list
2404 // to determine order.
2405 Node* SuperWord::executed_first(Node_List* p) {
2406 Node* n = p->at(0);
2407 int n_rpo = bb_idx(n);
2408 for (uint i = 1; i < p->size(); i++) {
2409 Node* s = p->at(i);
2410 int s_rpo = bb_idx(s);
2411 if (s_rpo < n_rpo) {
2412 n = s;
2413 n_rpo = s_rpo;
2414 }
2415 }
2416 return n;
2417 }
2418
2419 //------------------------------executed_last---------------------------
2420 // Return the node executed last in pack p.
2421 Node* SuperWord::executed_last(Node_List* p) {
2422 Node* n = p->at(0);
2423 int n_rpo = bb_idx(n);
2424 for (uint i = 1; i < p->size(); i++) {
2425 Node* s = p->at(i);
2426 int s_rpo = bb_idx(s);
2427 if (s_rpo > n_rpo) {
2428 n = s;
2429 n_rpo = s_rpo;
2430 }
2431 }
2432 return n;
2433 }
2434
2435 //----------------------------align_initial_loop_index---------------------------
2436 // Adjust pre-loop limit so that in main loop, a load/store reference
2437 // to align_to_ref will be a position zero in the vector.
2438 // (iv + k) mod vector_align == 0
2439 void SuperWord::align_initial_loop_index(MemNode* align_to_ref) {
2440 CountedLoopNode *main_head = lp()->as_CountedLoop();
2441 assert(main_head->is_main_loop(), "");
2442 CountedLoopEndNode* pre_end = get_pre_loop_end(main_head);
2443 assert(pre_end != NULL, "we must have a correct pre-loop");
2444 Node *pre_opaq1 = pre_end->limit();
2445 assert(pre_opaq1->Opcode() == Op_Opaque1, "");
2446 Opaque1Node *pre_opaq = (Opaque1Node*)pre_opaq1;
2447 Node *lim0 = pre_opaq->in(1);
2448
2449 // Where we put new limit calculations
2450 Node *pre_ctrl = pre_end->loopnode()->in(LoopNode::EntryControl);
2451
2452 // Ensure the original loop limit is available from the
2453 // pre-loop Opaque1 node.
2454 Node *orig_limit = pre_opaq->original_loop_limit();
2455 assert(orig_limit != NULL && _igvn.type(orig_limit) != Type::TOP, "");
2456
2457 SWPointer align_to_ref_p(align_to_ref, this, NULL, false);
2458 assert(align_to_ref_p.valid(), "sanity");
2459
2460 // Given:
2461 // lim0 == original pre loop limit
2462 // V == v_align (power of 2)
2463 // invar == extra invariant piece of the address expression
2464 // e == offset [ +/- invar ]
2465 //
2466 // When reassociating expressions involving '%' the basic rules are:
2467 // (a - b) % k == 0 => a % k == b % k
2468 // and:
2469 // (a + b) % k == 0 => a % k == (k - b) % k
2470 //
2471 // For stride > 0 && scale > 0,
2472 // Derive the new pre-loop limit "lim" such that the two constraints:
2473 // (1) lim = lim0 + N (where N is some positive integer < V)
2474 // (2) (e + lim) % V == 0
2475 // are true.
2476 //
2477 // Substituting (1) into (2),
2478 // (e + lim0 + N) % V == 0
2479 // solve for N:
2480 // N = (V - (e + lim0)) % V
2481 // substitute back into (1), so that new limit
2482 // lim = lim0 + (V - (e + lim0)) % V
2483 //
2484 // For stride > 0 && scale < 0
2485 // Constraints:
2486 // lim = lim0 + N
2487 // (e - lim) % V == 0
2488 // Solving for lim:
2489 // (e - lim0 - N) % V == 0
2490 // N = (e - lim0) % V
2491 // lim = lim0 + (e - lim0) % V
2492 //
2493 // For stride < 0 && scale > 0
2494 // Constraints:
2495 // lim = lim0 - N
2496 // (e + lim) % V == 0
2497 // Solving for lim:
2498 // (e + lim0 - N) % V == 0
2499 // N = (e + lim0) % V
2500 // lim = lim0 - (e + lim0) % V
2501 //
2502 // For stride < 0 && scale < 0
2503 // Constraints:
2504 // lim = lim0 - N
2505 // (e - lim) % V == 0
2506 // Solving for lim:
2507 // (e - lim0 + N) % V == 0
2508 // N = (V - (e - lim0)) % V
2509 // lim = lim0 - (V - (e - lim0)) % V
2510
2511 int vw = vector_width_in_bytes(align_to_ref);
2512 int stride = iv_stride();
2513 int scale = align_to_ref_p.scale_in_bytes();
2514 int elt_size = align_to_ref_p.memory_size();
2515 int v_align = vw / elt_size;
2516 assert(v_align > 1, "sanity");
2517 int offset = align_to_ref_p.offset_in_bytes() / elt_size;
2518 Node *offsn = _igvn.intcon(offset);
2519
2520 Node *e = offsn;
2521 if (align_to_ref_p.invar() != NULL) {
2522 // incorporate any extra invariant piece producing (offset +/- invar) >>> log2(elt)
2523 Node* log2_elt = _igvn.intcon(exact_log2(elt_size));
2524 Node* aref = new URShiftINode(align_to_ref_p.invar(), log2_elt);
2525 _igvn.register_new_node_with_optimizer(aref);
2526 _phase->set_ctrl(aref, pre_ctrl);
2527 if (align_to_ref_p.negate_invar()) {
2528 e = new SubINode(e, aref);
2529 } else {
2530 e = new AddINode(e, aref);
2531 }
2532 _igvn.register_new_node_with_optimizer(e);
2533 _phase->set_ctrl(e, pre_ctrl);
2534 }
2535 if (vw > ObjectAlignmentInBytes) {
2536 // incorporate base e +/- base && Mask >>> log2(elt)
2537 Node* xbase = new CastP2XNode(NULL, align_to_ref_p.base());
2538 _igvn.register_new_node_with_optimizer(xbase);
2539 #ifdef _LP64
2540 xbase = new ConvL2INode(xbase);
2541 _igvn.register_new_node_with_optimizer(xbase);
2542 #endif
2543 Node* mask = _igvn.intcon(vw-1);
2544 Node* masked_xbase = new AndINode(xbase, mask);
2545 _igvn.register_new_node_with_optimizer(masked_xbase);
2546 Node* log2_elt = _igvn.intcon(exact_log2(elt_size));
2547 Node* bref = new URShiftINode(masked_xbase, log2_elt);
2548 _igvn.register_new_node_with_optimizer(bref);
2549 _phase->set_ctrl(bref, pre_ctrl);
2550 e = new AddINode(e, bref);
2551 _igvn.register_new_node_with_optimizer(e);
2552 _phase->set_ctrl(e, pre_ctrl);
2553 }
2554
2555 // compute e +/- lim0
2556 if (scale < 0) {
2557 e = new SubINode(e, lim0);
2558 } else {
2559 e = new AddINode(e, lim0);
2560 }
2561 _igvn.register_new_node_with_optimizer(e);
2562 _phase->set_ctrl(e, pre_ctrl);
2563
2564 if (stride * scale > 0) {
2565 // compute V - (e +/- lim0)
2566 Node* va = _igvn.intcon(v_align);
2567 e = new SubINode(va, e);
2568 _igvn.register_new_node_with_optimizer(e);
2569 _phase->set_ctrl(e, pre_ctrl);
2570 }
2571 // compute N = (exp) % V
2572 Node* va_msk = _igvn.intcon(v_align - 1);
2573 Node* N = new AndINode(e, va_msk);
2574 _igvn.register_new_node_with_optimizer(N);
2575 _phase->set_ctrl(N, pre_ctrl);
2576
2577 // substitute back into (1), so that new limit
2578 // lim = lim0 + N
2579 Node* lim;
2580 if (stride < 0) {
2581 lim = new SubINode(lim0, N);
2582 } else {
2583 lim = new AddINode(lim0, N);
2584 }
2585 _igvn.register_new_node_with_optimizer(lim);
2586 _phase->set_ctrl(lim, pre_ctrl);
2587 Node* constrained =
2588 (stride > 0) ? (Node*) new MinINode(lim, orig_limit)
2589 : (Node*) new MaxINode(lim, orig_limit);
2590 _igvn.register_new_node_with_optimizer(constrained);
2591 _phase->set_ctrl(constrained, pre_ctrl);
2592 _igvn.hash_delete(pre_opaq);
2593 pre_opaq->set_req(1, constrained);
2594 }
2595
2596 //----------------------------get_pre_loop_end---------------------------
2597 // Find pre loop end from main loop. Returns null if none.
2598 CountedLoopEndNode* SuperWord::get_pre_loop_end(CountedLoopNode *cl) {
2599 Node *ctrl = cl->in(LoopNode::EntryControl);
2600 if (!ctrl->is_IfTrue() && !ctrl->is_IfFalse()) return NULL;
2601 Node *iffm = ctrl->in(0);
2602 if (!iffm->is_If()) return NULL;
2603 Node *p_f = iffm->in(0);
2604 if (!p_f->is_IfFalse()) return NULL;
2605 if (!p_f->in(0)->is_CountedLoopEnd()) return NULL;
2606 CountedLoopEndNode *pre_end = p_f->in(0)->as_CountedLoopEnd();
2607 CountedLoopNode* loop_node = pre_end->loopnode();
2608 if (loop_node == NULL || !loop_node->is_pre_loop()) return NULL;
2609 return pre_end;
2610 }
2611
2612
2613 //------------------------------init---------------------------
2614 void SuperWord::init() {
2615 _dg.init();
2616 _packset.clear();
2617 _disjoint_ptrs.clear();
2618 _block.clear();
2619 _data_entry.clear();
2620 _mem_slice_head.clear();
2621 _mem_slice_tail.clear();
2622 _iteration_first.clear();
2623 _iteration_last.clear();
2624 _node_info.clear();
2625 _align_to_ref = NULL;
2626 _lpt = NULL;
2627 _lp = NULL;
2628 _bb = NULL;
2629 _iv = NULL;
2630 _race_possible = 0;
2631 _early_return = false;
2632 _num_work_vecs = 0;
2633 _num_reductions = 0;
2634 }
2635
2636 //------------------------------restart---------------------------
2637 void SuperWord::restart() {
2638 _dg.init();
2639 _packset.clear();
2640 _disjoint_ptrs.clear();
2641 _block.clear();
2642 _data_entry.clear();
2643 _mem_slice_head.clear();
2644 _mem_slice_tail.clear();
2645 _node_info.clear();
2646 }
2647
2648 //------------------------------print_packset---------------------------
2649 void SuperWord::print_packset() {
2650 #ifndef PRODUCT
2651 tty->print_cr("packset");
2652 for (int i = 0; i < _packset.length(); i++) {
2653 tty->print_cr("Pack: %d", i);
2654 Node_List* p = _packset.at(i);
2655 print_pack(p);
2656 }
2657 #endif
2658 }
2659
2660 //------------------------------print_pack---------------------------
2661 void SuperWord::print_pack(Node_List* p) {
2662 for (uint i = 0; i < p->size(); i++) {
2663 print_stmt(p->at(i));
2664 }
2665 }
2666
2667 //------------------------------print_bb---------------------------
2668 void SuperWord::print_bb() {
2669 #ifndef PRODUCT
2670 tty->print_cr("\nBlock");
2671 for (int i = 0; i < _block.length(); i++) {
2672 Node* n = _block.at(i);
2673 tty->print("%d ", i);
2674 if (n) {
2675 n->dump();
2676 }
2677 }
2678 #endif
2679 }
2680
2681 //------------------------------print_stmt---------------------------
2682 void SuperWord::print_stmt(Node* s) {
2683 #ifndef PRODUCT
2684 tty->print(" align: %d \t", alignment(s));
2685 s->dump();
2686 #endif
2687 }
2688
2689 //------------------------------blank---------------------------
2690 char* SuperWord::blank(uint depth) {
2691 static char blanks[101];
2692 assert(depth < 101, "too deep");
2693 for (uint i = 0; i < depth; i++) blanks[i] = ' ';
2694 blanks[depth] = '\0';
2695 return blanks;
2696 }
2697
2698
2699 //==============================SWPointer===========================
2700
2701 //----------------------------SWPointer------------------------
2702 SWPointer::SWPointer(MemNode* mem, SuperWord* slp, Node_Stack *nstack, bool analyze_only) :
2703 _mem(mem), _slp(slp), _base(NULL), _adr(NULL),
2704 _scale(0), _offset(0), _invar(NULL), _negate_invar(false),
2705 _nstack(nstack), _analyze_only(analyze_only),
2706 _stack_idx(0) {
2707
2708 Node* adr = mem->in(MemNode::Address);
2709 if (!adr->is_AddP()) {
2710 assert(!valid(), "too complex");
2711 return;
2712 }
2713 // Match AddP(base, AddP(ptr, k*iv [+ invariant]), constant)
2714 Node* base = adr->in(AddPNode::Base);
2715 // The base address should be loop invariant
2716 if (!invariant(base)) {
2717 assert(!valid(), "base address is loop variant");
2718 return;
2719 }
2720 //unsafe reference could not be aligned appropriately without runtime checking
2721 if (base == NULL || base->bottom_type() == Type::TOP) {
2722 assert(!valid(), "unsafe access");
2723 return;
2724 }
2725 for (int i = 0; i < 3; i++) {
2726 if (!scaled_iv_plus_offset(adr->in(AddPNode::Offset))) {
2727 assert(!valid(), "too complex");
2728 return;
2729 }
2730 adr = adr->in(AddPNode::Address);
2731 if (base == adr || !adr->is_AddP()) {
2732 break; // stop looking at addp's
2733 }
2734 }
2735 _base = base;
2736 _adr = adr;
2737 assert(valid(), "Usable");
2738 }
2739
2740 // Following is used to create a temporary object during
2741 // the pattern match of an address expression.
2742 SWPointer::SWPointer(SWPointer* p) :
2743 _mem(p->_mem), _slp(p->_slp), _base(NULL), _adr(NULL),
2744 _scale(0), _offset(0), _invar(NULL), _negate_invar(false),
2745 _nstack(p->_nstack), _analyze_only(p->_analyze_only),
2746 _stack_idx(p->_stack_idx) {}
2747
2748 //------------------------scaled_iv_plus_offset--------------------
2749 // Match: k*iv + offset
2750 // where: k is a constant that maybe zero, and
2751 // offset is (k2 [+/- invariant]) where k2 maybe zero and invariant is optional
2752 bool SWPointer::scaled_iv_plus_offset(Node* n) {
2753 if (scaled_iv(n)) {
2754 return true;
2755 }
2756 if (offset_plus_k(n)) {
2757 return true;
2758 }
2759 int opc = n->Opcode();
2760 if (opc == Op_AddI) {
2761 if (scaled_iv(n->in(1)) && offset_plus_k(n->in(2))) {
2762 return true;
2763 }
2764 if (scaled_iv(n->in(2)) && offset_plus_k(n->in(1))) {
2765 return true;
2766 }
2767 } else if (opc == Op_SubI) {
2768 if (scaled_iv(n->in(1)) && offset_plus_k(n->in(2), true)) {
2769 return true;
2770 }
2771 if (scaled_iv(n->in(2)) && offset_plus_k(n->in(1))) {
2772 _scale *= -1;
2773 return true;
2774 }
2775 }
2776 return false;
2777 }
2778
2779 //----------------------------scaled_iv------------------------
2780 // Match: k*iv where k is a constant that's not zero
2781 bool SWPointer::scaled_iv(Node* n) {
2782 if (_scale != 0) {
2783 return false; // already found a scale
2784 }
2785 if (n == iv()) {
2786 _scale = 1;
2787 return true;
2788 }
2789 if (_analyze_only && (invariant(n) == false)) {
2790 _nstack->push(n, _stack_idx++);
2791 }
2792 int opc = n->Opcode();
2793 if (opc == Op_MulI) {
2794 if (n->in(1) == iv() && n->in(2)->is_Con()) {
2795 _scale = n->in(2)->get_int();
2796 return true;
2797 } else if (n->in(2) == iv() && n->in(1)->is_Con()) {
2798 _scale = n->in(1)->get_int();
2799 return true;
2800 }
2801 } else if (opc == Op_LShiftI) {
2802 if (n->in(1) == iv() && n->in(2)->is_Con()) {
2803 _scale = 1 << n->in(2)->get_int();
2804 return true;
2805 }
2806 } else if (opc == Op_ConvI2L) {
2807 if (scaled_iv_plus_offset(n->in(1))) {
2808 return true;
2809 }
2810 } else if (opc == Op_LShiftL) {
2811 if (!has_iv() && _invar == NULL) {
2812 // Need to preserve the current _offset value, so
2813 // create a temporary object for this expression subtree.
2814 // Hacky, so should re-engineer the address pattern match.
2815 SWPointer tmp(this);
2816 if (tmp.scaled_iv_plus_offset(n->in(1))) {
2817 if (tmp._invar == NULL) {
2818 int mult = 1 << n->in(2)->get_int();
2819 _scale = tmp._scale * mult;
2820 _offset += tmp._offset * mult;
2821 return true;
2822 }
2823 }
2824 }
2825 }
2826 return false;
2827 }
2828
2829 //----------------------------offset_plus_k------------------------
2830 // Match: offset is (k [+/- invariant])
2831 // where k maybe zero and invariant is optional, but not both.
2832 bool SWPointer::offset_plus_k(Node* n, bool negate) {
2833 int opc = n->Opcode();
2834 if (opc == Op_ConI) {
2835 _offset += negate ? -(n->get_int()) : n->get_int();
2836 return true;
2837 } else if (opc == Op_ConL) {
2838 // Okay if value fits into an int
2839 const TypeLong* t = n->find_long_type();
2840 if (t->higher_equal(TypeLong::INT)) {
2841 jlong loff = n->get_long();
2842 jint off = (jint)loff;
2843 _offset += negate ? -off : loff;
2844 return true;
2845 }
2846 return false;
2847 }
2848 if (_invar != NULL) return false; // already have an invariant
2849 if (_analyze_only && (invariant(n) == false)) {
2850 _nstack->push(n, _stack_idx++);
2851 }
2852 if (opc == Op_AddI) {
2853 if (n->in(2)->is_Con() && invariant(n->in(1))) {
2854 _negate_invar = negate;
2855 _invar = n->in(1);
2856 _offset += negate ? -(n->in(2)->get_int()) : n->in(2)->get_int();
2857 return true;
2858 } else if (n->in(1)->is_Con() && invariant(n->in(2))) {
2859 _offset += negate ? -(n->in(1)->get_int()) : n->in(1)->get_int();
2860 _negate_invar = negate;
2861 _invar = n->in(2);
2862 return true;
2863 }
2864 }
2865 if (opc == Op_SubI) {
2866 if (n->in(2)->is_Con() && invariant(n->in(1))) {
2867 _negate_invar = negate;
2868 _invar = n->in(1);
2869 _offset += !negate ? -(n->in(2)->get_int()) : n->in(2)->get_int();
2870 return true;
2871 } else if (n->in(1)->is_Con() && invariant(n->in(2))) {
2872 _offset += negate ? -(n->in(1)->get_int()) : n->in(1)->get_int();
2873 _negate_invar = !negate;
2874 _invar = n->in(2);
2875 return true;
2876 }
2877 }
2878 if (invariant(n)) {
2879 _negate_invar = negate;
2880 _invar = n;
2881 return true;
2882 }
2883 return false;
2884 }
2885
2886 //----------------------------print------------------------
2887 void SWPointer::print() {
2888 #ifndef PRODUCT
2889 tty->print("base: %d adr: %d scale: %d offset: %d invar: %c%d\n",
2890 _base != NULL ? _base->_idx : 0,
2891 _adr != NULL ? _adr->_idx : 0,
2892 _scale, _offset,
2893 _negate_invar?'-':'+',
2894 _invar != NULL ? _invar->_idx : 0);
2895 #endif
2896 }
2897
2898 // ========================= OrderedPair =====================
2899
2900 const OrderedPair OrderedPair::initial;
2901
2902 // ========================= SWNodeInfo =====================
2903
2904 const SWNodeInfo SWNodeInfo::initial;
2905
2906
2907 // ============================ DepGraph ===========================
2908
2909 //------------------------------make_node---------------------------
2910 // Make a new dependence graph node for an ideal node.
2911 DepMem* DepGraph::make_node(Node* node) {
2912 DepMem* m = new (_arena) DepMem(node);
2913 if (node != NULL) {
2914 assert(_map.at_grow(node->_idx) == NULL, "one init only");
2915 _map.at_put_grow(node->_idx, m);
2916 }
2917 return m;
2918 }
2919
2920 //------------------------------make_edge---------------------------
2921 // Make a new dependence graph edge from dpred -> dsucc
2922 DepEdge* DepGraph::make_edge(DepMem* dpred, DepMem* dsucc) {
2923 DepEdge* e = new (_arena) DepEdge(dpred, dsucc, dsucc->in_head(), dpred->out_head());
2924 dpred->set_out_head(e);
2925 dsucc->set_in_head(e);
2926 return e;
2927 }
2928
2929 // ========================== DepMem ========================
2930
2931 //------------------------------in_cnt---------------------------
2932 int DepMem::in_cnt() {
2933 int ct = 0;
2934 for (DepEdge* e = _in_head; e != NULL; e = e->next_in()) ct++;
2935 return ct;
2936 }
2937
2938 //------------------------------out_cnt---------------------------
2939 int DepMem::out_cnt() {
2940 int ct = 0;
2941 for (DepEdge* e = _out_head; e != NULL; e = e->next_out()) ct++;
2942 return ct;
2943 }
2944
2945 //------------------------------print-----------------------------
2946 void DepMem::print() {
2947 #ifndef PRODUCT
2948 tty->print(" DepNode %d (", _node->_idx);
2949 for (DepEdge* p = _in_head; p != NULL; p = p->next_in()) {
2950 Node* pred = p->pred()->node();
2951 tty->print(" %d", pred != NULL ? pred->_idx : 0);
2952 }
2953 tty->print(") [");
2954 for (DepEdge* s = _out_head; s != NULL; s = s->next_out()) {
2955 Node* succ = s->succ()->node();
2956 tty->print(" %d", succ != NULL ? succ->_idx : 0);
2957 }
2958 tty->print_cr(" ]");
2959 #endif
2960 }
2961
2962 // =========================== DepEdge =========================
2963
2964 //------------------------------DepPreds---------------------------
2965 void DepEdge::print() {
2966 #ifndef PRODUCT
2967 tty->print_cr("DepEdge: %d [ %d ]", _pred->node()->_idx, _succ->node()->_idx);
2968 #endif
2969 }
2970
2971 // =========================== DepPreds =========================
2972 // Iterator over predecessor edges in the dependence graph.
2973
2974 //------------------------------DepPreds---------------------------
2975 DepPreds::DepPreds(Node* n, DepGraph& dg) {
2976 _n = n;
2977 _done = false;
2978 if (_n->is_Store() || _n->is_Load()) {
2979 _next_idx = MemNode::Address;
2980 _end_idx = n->req();
2981 _dep_next = dg.dep(_n)->in_head();
2982 } else if (_n->is_Mem()) {
2983 _next_idx = 0;
2984 _end_idx = 0;
2985 _dep_next = dg.dep(_n)->in_head();
2986 } else {
2987 _next_idx = 1;
2988 _end_idx = _n->req();
2989 _dep_next = NULL;
2990 }
2991 next();
2992 }
2993
2994 //------------------------------next---------------------------
2995 void DepPreds::next() {
2996 if (_dep_next != NULL) {
2997 _current = _dep_next->pred()->node();
2998 _dep_next = _dep_next->next_in();
2999 } else if (_next_idx < _end_idx) {
3000 _current = _n->in(_next_idx++);
3001 } else {
3002 _done = true;
3003 }
3004 }
3005
3006 // =========================== DepSuccs =========================
3007 // Iterator over successor edges in the dependence graph.
3008
3009 //------------------------------DepSuccs---------------------------
3010 DepSuccs::DepSuccs(Node* n, DepGraph& dg) {
3011 _n = n;
3012 _done = false;
3013 if (_n->is_Load()) {
3014 _next_idx = 0;
3015 _end_idx = _n->outcnt();
3016 _dep_next = dg.dep(_n)->out_head();
3017 } else if (_n->is_Mem() || _n->is_Phi() && _n->bottom_type() == Type::MEMORY) {
3018 _next_idx = 0;
3019 _end_idx = 0;
3020 _dep_next = dg.dep(_n)->out_head();
3021 } else {
3022 _next_idx = 0;
3023 _end_idx = _n->outcnt();
3024 _dep_next = NULL;
3025 }
3026 next();
3027 }
3028
3029 //-------------------------------next---------------------------
3030 void DepSuccs::next() {
3031 if (_dep_next != NULL) {
3032 _current = _dep_next->succ()->node();
3033 _dep_next = _dep_next->next_out();
3034 } else if (_next_idx < _end_idx) {
3035 _current = _n->raw_out(_next_idx++);
3036 } else {
3037 _done = true;
3038 }
3039 }
3040
3041 //
3042 // --------------------------------- vectorization/simd -----------------------------------
3043 //
3044 Node* SuperWord::find_phi_for_mem_dep(LoadNode* ld) {
3045 assert(in_bb(ld), "must be in block");
3046 if (_clone_map.gen(ld->_idx) == _ii_first) {
3047 #ifndef PRODUCT
3048 if (_vector_loop_debug) {
3049 tty->print_cr("SuperWord::find_phi_for_mem_dep _clone_map.gen(ld->_idx)=%d",
3050 _clone_map.gen(ld->_idx));
3051 }
3052 #endif
3053 return NULL; //we think that any ld in the first gen being vectorizable
3054 }
3055
3056 Node* mem = ld->in(MemNode::Memory);
3057 if (mem->outcnt() <= 1) {
3058 // we don't want to remove the only edge from mem node to load
3059 #ifndef PRODUCT
3060 if (_vector_loop_debug) {
3061 tty->print_cr("SuperWord::find_phi_for_mem_dep input node %d to load %d has no other outputs and edge mem->load cannot be removed",
3062 mem->_idx, ld->_idx);
3063 ld->dump();
3064 mem->dump();
3065 }
3066 #endif
3067 return NULL;
3068 }
3069 if (!in_bb(mem) || _clone_map.gen(mem->_idx) == _clone_map.gen(ld->_idx)) {
3070 #ifndef PRODUCT
3071 if (_vector_loop_debug) {
3072 tty->print_cr("SuperWord::find_phi_for_mem_dep _clone_map.gen(mem->_idx)=%d",
3073 _clone_map.gen(mem->_idx));
3074 }
3075 #endif
3076 return NULL; // does not depend on loop volatile node or depends on the same generation
3077 }
3078
3079 //otherwise first node should depend on mem-phi
3080 Node* first = first_node(ld);
3081 assert(first->is_Load(), "must be Load");
3082 Node* phi = first->as_Load()->in(MemNode::Memory);
3083 if (!phi->is_Phi() || phi->bottom_type() != Type::MEMORY) {
3084 #ifndef PRODUCT
3085 if (_vector_loop_debug) {
3086 tty->print_cr("SuperWord::find_phi_for_mem_dep load is not vectorizable node, since it's `first` does not take input from mem phi");
3087 ld->dump();
3088 first->dump();
3089 }
3090 #endif
3091 return NULL;
3092 }
3093
3094 Node* tail = 0;
3095 for (int m = 0; m < _mem_slice_head.length(); m++) {
3096 if (_mem_slice_head.at(m) == phi) {
3097 tail = _mem_slice_tail.at(m);
3098 }
3099 }
3100 if (tail == 0) { //test that found phi is in the list _mem_slice_head
3101 #ifndef PRODUCT
3102 if (_vector_loop_debug) {
3103 tty->print_cr("SuperWord::find_phi_for_mem_dep load %d is not vectorizable node, its phi %d is not _mem_slice_head",
3104 ld->_idx, phi->_idx);
3105 ld->dump();
3106 phi->dump();
3107 }
3108 #endif
3109 return NULL;
3110 }
3111
3112 // now all conditions are met
3113 return phi;
3114 }
3115
3116 Node* SuperWord::first_node(Node* nd) {
3117 for (int ii = 0; ii < _iteration_first.length(); ii++) {
3118 Node* nnn = _iteration_first.at(ii);
3119 if (_clone_map.idx(nnn->_idx) == _clone_map.idx(nd->_idx)) {
3120 #ifndef PRODUCT
3121 if (_vector_loop_debug) {
3122 tty->print_cr("SuperWord::first_node: %d is the first iteration node for %d (_clone_map.idx(nnn->_idx) = %d)",
3123 nnn->_idx, nd->_idx, _clone_map.idx(nnn->_idx));
3124 }
3125 #endif
3126 return nnn;
3127 }
3128 }
3129
3130 #ifndef PRODUCT
3131 if (_vector_loop_debug) {
3132 tty->print_cr("SuperWord::first_node: did not find first iteration node for %d (_clone_map.idx(nd->_idx)=%d)",
3133 nd->_idx, _clone_map.idx(nd->_idx));
3134 }
3135 #endif
3136 return 0;
3137 }
3138
3139 Node* SuperWord::last_node(Node* nd) {
3140 for (int ii = 0; ii < _iteration_last.length(); ii++) {
3141 Node* nnn = _iteration_last.at(ii);
3142 if (_clone_map.idx(nnn->_idx) == _clone_map.idx(nd->_idx)) {
3143 #ifndef PRODUCT
3144 if (_vector_loop_debug) {
3145 tty->print_cr("SuperWord::last_node _clone_map.idx(nnn->_idx)=%d, _clone_map.idx(nd->_idx)=%d",
3146 _clone_map.idx(nnn->_idx), _clone_map.idx(nd->_idx));
3147 }
3148 #endif
3149 return nnn;
3150 }
3151 }
3152 return 0;
3153 }
3154
3155 int SuperWord::mark_generations() {
3156 Node *ii_err = 0, *tail_err;
3157 for (int i = 0; i < _mem_slice_head.length(); i++) {
3158 Node* phi = _mem_slice_head.at(i);
3159 assert(phi->is_Phi(), "must be phi");
3160
3161 Node* tail = _mem_slice_tail.at(i);
3162 if (_ii_last == -1) {
3163 tail_err = tail;
3164 _ii_last = _clone_map.gen(tail->_idx);
3165 }
3166 else if (_ii_last != _clone_map.gen(tail->_idx)) {
3167 #ifndef PRODUCT
3168 if (TraceSuperWord && Verbose) {
3169 tty->print_cr("SuperWord::mark_generations _ii_last error - found different generations in two tail nodes ");
3170 tail->dump();
3171 tail_err->dump();
3172 }
3173 #endif
3174 return -1;
3175 }
3176
3177 // find first iteration in the loop
3178 for (DUIterator_Fast imax, i = phi->fast_outs(imax); i < imax; i++) {
3179 Node* ii = phi->fast_out(i);
3180 if (in_bb(ii) && ii->is_Store()) { // we speculate that normally Stores of one and one only generation have deps from mem phi
3181 if (_ii_first == -1) {
3182 ii_err = ii;
3183 _ii_first = _clone_map.gen(ii->_idx);
3184 } else if (_ii_first != _clone_map.gen(ii->_idx)) {
3185 #ifndef PRODUCT
3186 if (TraceSuperWord && Verbose) {
3187 tty->print_cr("SuperWord::mark_generations _ii_first error - found different generations in two nodes ");
3188 ii->dump();
3189 ii_err->dump();
3190 }
3191 #endif
3192 return -1; // this phi has Stores from different generations of unroll and cannot be simd/vectorized
3193 }
3194 }
3195 }//for (DUIterator_Fast imax,
3196 }//for (int i...
3197
3198 if (_ii_first == -1 || _ii_last == -1) {
3199 #ifndef PRODUCT
3200 if (TraceSuperWord && Verbose) {
3201 tty->print_cr("SuperWord::mark_generations unknown error, something vent wrong");
3202 }
3203 #endif
3204 return -1; // something vent wrong
3205 }
3206 // collect nodes in the first and last generations
3207 assert(_iteration_first.length() == 0, "_iteration_first must be empty");
3208 assert(_iteration_last.length() == 0, "_iteration_last must be empty");
3209 for (int j = 0; j < _block.length(); j++) {
3210 Node* n = _block.at(j);
3211 node_idx_t gen = _clone_map.gen(n->_idx);
3212 if ((signed)gen == _ii_first) {
3213 _iteration_first.push(n);
3214 } else if ((signed)gen == _ii_last) {
3215 _iteration_last.push(n);
3216 }
3217 }
3218
3219 // building order of iterations
3220 assert(_ii_order.length() == 0, "should be empty");
3221 if (ii_err != 0) {
3222 assert(in_bb(ii_err) && ii_err->is_Store(), "should be Store in bb");
3223 Node* nd = ii_err;
3224 while(_clone_map.gen(nd->_idx) != _ii_last) {
3225 _ii_order.push(_clone_map.gen(nd->_idx));
3226 bool found = false;
3227 for (DUIterator_Fast imax, i = nd->fast_outs(imax); i < imax; i++) {
3228 Node* use = nd->fast_out(i);
3229 if (_clone_map.idx(use->_idx) == _clone_map.idx(nd->_idx) && use->as_Store()->in(MemNode::Memory) == nd) {
3230 found = true;
3231 nd = use;
3232 break;
3233 }
3234 }//for
3235
3236 if (found == false) {
3237 #ifndef PRODUCT
3238 if (TraceSuperWord && Verbose) {
3239 tty->print_cr("SuperWord::mark_generations: Cannot build order of iterations - no dependent Store for %d", nd->_idx);
3240 }
3241 #endif
3242 _ii_order.clear();
3243 return -1;
3244 }
3245 } //while
3246 _ii_order.push(_clone_map.gen(nd->_idx));
3247 }
3248
3249 #ifndef PRODUCT
3250 if (_vector_loop_debug) {
3251 tty->print_cr("SuperWord::mark_generations");
3252 tty->print_cr("First generation (%d) nodes:", _ii_first);
3253 for (int ii = 0; ii < _iteration_first.length(); ii++) _iteration_first.at(ii)->dump();
3254 tty->print_cr("Last generation (%d) nodes:", _ii_last);
3255 for (int ii = 0; ii < _iteration_last.length(); ii++) _iteration_last.at(ii)->dump();
3256 tty->print_cr(" ");
3257
3258 tty->print("SuperWord::List of generations: ");
3259 for (int jj = 0; jj < _ii_order.length(); ++jj) {
3260 tty->print("%d:%d ", jj, _ii_order.at(jj));
3261 }
3262 tty->print_cr(" ");
3263 }
3264 #endif
3265
3266 return _ii_first;
3267 }
3268
3269 bool SuperWord::fix_commutative_inputs(Node* gold, Node* fix) {
3270 assert(gold->is_Add() && fix->is_Add() || gold->is_Mul() && fix->is_Mul(), "should be only Add or Mul nodes");
3271 assert(_clone_map.idx(gold->_idx) == _clone_map.idx(fix->_idx), "should be clones of the same node");
3272 Node* gin1 = gold->in(1);
3273 Node* gin2 = gold->in(2);
3274 Node* fin1 = fix->in(1);
3275 Node* fin2 = fix->in(2);
3276 bool swapped = false;
3277
3278 if (in_bb(gin1) && in_bb(gin2) && in_bb(fin1) && in_bb(fin1)) {
3279 if (_clone_map.idx(gin1->_idx) == _clone_map.idx(fin1->_idx) &&
3280 _clone_map.idx(gin2->_idx) == _clone_map.idx(fin2->_idx)) {
3281 return true; // nothing to fix
3282 }
3283 if (_clone_map.idx(gin1->_idx) == _clone_map.idx(fin2->_idx) &&
3284 _clone_map.idx(gin2->_idx) == _clone_map.idx(fin1->_idx)) {
3285 fix->swap_edges(1, 2);
3286 swapped = true;
3287 }
3288 }
3289 // at least one input comes from outside of bb
3290 if (gin1->_idx == fin1->_idx) {
3291 return true; // nothing to fix
3292 }
3293 if (!swapped && (gin1->_idx == fin2->_idx || gin2->_idx == fin1->_idx)) { //swapping is expensive, check condition first
3294 fix->swap_edges(1, 2);
3295 swapped = true;
3296 }
3297
3298 if (swapped) {
3299 #ifndef PRODUCT
3300 if (_vector_loop_debug) {
3301 tty->print_cr("SuperWord::fix_commutative_inputs: fixed node %d", fix->_idx);
3302 }
3303 #endif
3304 return true;
3305 }
3306
3307 #ifndef PRODUCT
3308 if (TraceSuperWord && Verbose) {
3309 tty->print_cr("SuperWord::fix_commutative_inputs: cannot fix node %d", fix->_idx);
3310 }
3311 #endif
3312 return false;
3313 }
3314
3315 bool SuperWord::pack_parallel() {
3316 #ifndef PRODUCT
3317 if (_vector_loop_debug) {
3318 tty->print_cr("SuperWord::pack_parallel: START");
3319 }
3320 #endif
3321
3322 _packset.clear();
3323
3324 for (int ii = 0; ii < _iteration_first.length(); ii++) {
3325 Node* nd = _iteration_first.at(ii);
3326 if (in_bb(nd) && (nd->is_Load() || nd->is_Store() || nd->is_Add() || nd->is_Mul())) {
3327 Node_List* pk = new Node_List();
3328 pk->push(nd);
3329 for (int gen = 1; gen < _ii_order.length(); ++gen) {
3330 for (int kk = 0; kk < _block.length(); kk++) {
3331 Node* clone = _block.at(kk);
3332 if (_clone_map.idx(clone->_idx) == _clone_map.idx(nd->_idx) &&
3333 _clone_map.gen(clone->_idx) == _ii_order.at(gen)) {
3334 if (nd->is_Add() || nd->is_Mul()) {
3335 fix_commutative_inputs(nd, clone);
3336 }
3337 pk->push(clone);
3338 if (pk->size() == 4) {
3339 _packset.append(pk);
3340 #ifndef PRODUCT
3341 if (_vector_loop_debug) {
3342 tty->print_cr("SuperWord::pack_parallel: added pack ");
3343 pk->dump();
3344 }
3345 #endif
3346 if (_clone_map.gen(clone->_idx) != _ii_last) {
3347 pk = new Node_List();
3348 }
3349 }
3350 break;
3351 }
3352 }
3353 }//for
3354 }//if
3355 }//for
3356
3357 #ifndef PRODUCT
3358 if (_vector_loop_debug) {
3359 tty->print_cr("SuperWord::pack_parallel: END");
3360 }
3361 #endif
3362
3363 return true;
3364 }
3365
3366 bool SuperWord::hoist_loads_in_graph() {
3367 GrowableArray<Node*> loads;
3368
3369 #ifndef PRODUCT
3370 if (_vector_loop_debug) {
3371 tty->print_cr("SuperWord::hoist_loads_in_graph: total number _mem_slice_head.length() = %d", _mem_slice_head.length());
3372 }
3373 #endif
3374
3375 for (int i = 0; i < _mem_slice_head.length(); i++) {
3376 Node* n = _mem_slice_head.at(i);
3377 if ( !in_bb(n) || !n->is_Phi() || n->bottom_type() != Type::MEMORY) {
3378 #ifndef PRODUCT
3379 if (TraceSuperWord && Verbose) {
3380 tty->print_cr("SuperWord::hoist_loads_in_graph: skipping unexpected node n=%d", n->_idx);
3381 }
3382 #endif
3383 continue;
3384 }
3385
3386 #ifndef PRODUCT
3387 if (_vector_loop_debug) {
3388 tty->print_cr("SuperWord::hoist_loads_in_graph: processing phi %d = _mem_slice_head.at(%d);", n->_idx, i);
3389 }
3390 #endif
3391
3392 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3393 Node* ld = n->fast_out(i);
3394 if (ld->is_Load() && ld->as_Load()->in(MemNode::Memory) == n && in_bb(ld)) {
3395 for (int i = 0; i < _block.length(); i++) {
3396 Node* ld2 = _block.at(i);
3397 if (ld2->is_Load() &&
3398 _clone_map.idx(ld->_idx) == _clone_map.idx(ld2->_idx) &&
3399 _clone_map.gen(ld->_idx) != _clone_map.gen(ld2->_idx)) { // <= do not collect the first generation ld
3400 #ifndef PRODUCT
3401 if (_vector_loop_debug) {
3402 tty->print_cr("SuperWord::hoist_loads_in_graph: will try to hoist load ld2->_idx=%d, cloned from %d (ld->_idx=%d)",
3403 ld2->_idx, _clone_map.idx(ld->_idx), ld->_idx);
3404 }
3405 #endif
3406 // could not do on-the-fly, since iterator is immutable
3407 loads.push(ld2);
3408 }
3409 }// for
3410 }//if
3411 }//for (DUIterator_Fast imax,
3412 }//for (int i = 0; i
3413
3414 for (int i = 0; i < loads.length(); i++) {
3415 LoadNode* ld = loads.at(i)->as_Load();
3416 Node* phi = find_phi_for_mem_dep(ld);
3417 if (phi != NULL) {
3418 #ifndef PRODUCT
3419 if (_vector_loop_debug) {
3420 tty->print_cr("SuperWord::hoist_loads_in_graph replacing MemNode::Memory(%d) edge in %d with one from %d",
3421 MemNode::Memory, ld->_idx, phi->_idx);
3422 }
3423 #endif
3424 _igvn.replace_input_of(ld, MemNode::Memory, phi);
3425 }
3426 }//for
3427
3428 restart(); // invalidate all basic structures, since we rebuilt the graph
3429
3430 #ifndef PRODUCT
3431 if (TraceSuperWord && Verbose) {
3432 tty->print_cr("\nSuperWord::hoist_loads_in_graph() the graph was rebuilt, all structures invalidated and need rebuild");
3433 }
3434 #endif
3435 return true;
3436 }
3437