1 /*
2 * Copyright (c) 2007, 2016, 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 "memory/resourceArea.hpp"
29 #include "opto/addnode.hpp"
30 #include "opto/callnode.hpp"
31 #include "opto/castnode.hpp"
32 #include "opto/convertnode.hpp"
33 #include "opto/divnode.hpp"
34 #include "opto/matcher.hpp"
35 #include "opto/memnode.hpp"
36 #include "opto/mulnode.hpp"
37 #include "opto/opcodes.hpp"
38 #include "opto/opaquenode.hpp"
39 #include "opto/superword.hpp"
40 #include "opto/vectornode.hpp"
41 #include "opto/movenode.hpp"
42
43 //
44 // S U P E R W O R D T R A N S F O R M
45 //=============================================================================
46
47 //------------------------------SuperWord---------------------------
48 SuperWord::SuperWord(PhaseIdealLoop* phase) :
49 _phase(phase),
50 _igvn(phase->_igvn),
51 _arena(phase->C->comp_arena()),
52 _packset(arena(), 8, 0, NULL), // packs for the current block
53 _bb_idx(arena(), (int)(1.10 * phase->C->unique()), 0, 0), // node idx to index in bb
54 _block(arena(), 8, 0, NULL), // nodes in current block
55 _data_entry(arena(), 8, 0, NULL), // nodes with all inputs from outside
56 _mem_slice_head(arena(), 8, 0, NULL), // memory slice heads
57 _mem_slice_tail(arena(), 8, 0, NULL), // memory slice tails
58 _node_info(arena(), 8, 0, SWNodeInfo::initial), // info needed per node
59 _clone_map(phase->C->clone_map()), // map of nodes created in cloning
60 _cmovev_kit(_arena, this), // map to facilitate CMoveVD creation
61 _align_to_ref(NULL), // memory reference to align vectors to
62 _disjoint_ptrs(arena(), 8, 0, OrderedPair::initial), // runtime disambiguated pointer pairs
63 _dg(_arena), // dependence graph
64 _visited(arena()), // visited node set
65 _post_visited(arena()), // post visited node set
66 _n_idx_list(arena(), 8), // scratch list of (node,index) pairs
67 _stk(arena(), 8, 0, NULL), // scratch stack of nodes
68 _nlist(arena(), 8, 0, NULL), // scratch list of nodes
69 _lpt(NULL), // loop tree node
70 _lp(NULL), // LoopNode
71 _bb(NULL), // basic block
72 _iv(NULL), // induction var
73 _race_possible(false), // cases where SDMU is true
74 _early_return(true), // analysis evaluations routine
75 _num_work_vecs(0), // amount of vector work we have
76 _num_reductions(0), // amount of reduction work we have
77 _do_vector_loop(phase->C->do_vector_loop()), // whether to do vectorization/simd style
78 _do_reserve_copy(DoReserveCopyInSuperWord),
79 _ii_first(-1), // first loop generation index - only if do_vector_loop()
80 _ii_last(-1), // last loop generation index - only if do_vector_loop()
81 _ii_order(arena(), 8, 0, 0)
82 {
83 #ifndef PRODUCT
84 _vector_loop_debug = 0;
85 if (_phase->C->method() != NULL) {
86 _vector_loop_debug = phase->C->directive()->VectorizeDebugOption;
87 }
88
89 #endif
90 }
91
92 //------------------------------transform_loop---------------------------
93 void SuperWord::transform_loop(IdealLoopTree* lpt, bool do_optimization) {
94 assert(UseSuperWord, "should be");
95 // Do vectors exist on this architecture?
96 if (Matcher::vector_width_in_bytes(T_BYTE) < 2) return;
97
98 assert(lpt->_head->is_CountedLoop(), "must be");
99 CountedLoopNode *cl = lpt->_head->as_CountedLoop();
100
101 if (!cl->is_valid_counted_loop()) return; // skip malformed counted loop
102
103 if (!cl->is_main_loop() ) return; // skip normal, pre, and post loops
104 // Check for no control flow in body (other than exit)
105 Node *cl_exit = cl->loopexit();
106 if (cl_exit->in(0) != lpt->_head) {
107 #ifndef PRODUCT
108 if (TraceSuperWord) {
109 tty->print_cr("SuperWord::transform_loop: loop too complicated, cl_exit->in(0) != lpt->_head");
110 tty->print("cl_exit %d", cl_exit->_idx); cl_exit->dump();
111 tty->print("cl_exit->in(0) %d", cl_exit->in(0)->_idx); cl_exit->in(0)->dump();
112 tty->print("lpt->_head %d", lpt->_head->_idx); lpt->_head->dump();
113 lpt->dump_head();
114 }
115 #endif
116 return;
117 }
118
119 // Make sure the are no extra control users of the loop backedge
120 if (cl->back_control()->outcnt() != 1) {
121 return;
122 }
123
124 // We only re-enter slp when we vector mapped a queried loop and we want to
125 // continue unrolling, in this case, slp is not subsequently done.
126 if (cl->do_unroll_only()) return;
127
128 // Check for pre-loop ending with CountedLoopEnd(Bool(Cmp(x,Opaque1(limit))))
129 CountedLoopEndNode* pre_end = get_pre_loop_end(cl);
130 if (pre_end == NULL) return;
131 Node *pre_opaq1 = pre_end->limit();
132 if (pre_opaq1->Opcode() != Op_Opaque1) return;
133
134 init(); // initialize data structures
135
136 set_lpt(lpt);
137 set_lp(cl);
138
139 // For now, define one block which is the entire loop body
140 set_bb(cl);
141
142 if (do_optimization) {
143 assert(_packset.length() == 0, "packset must be empty");
144 SLP_extract();
145 }
146 }
147
148 //------------------------------early unrolling analysis------------------------------
149 void SuperWord::unrolling_analysis(int &local_loop_unroll_factor) {
150 bool is_slp = true;
151 ResourceMark rm;
152 size_t ignored_size = lpt()->_body.size();
153 int *ignored_loop_nodes = NEW_RESOURCE_ARRAY(int, ignored_size);
154 Node_Stack nstack((int)ignored_size);
155 CountedLoopNode *cl = lpt()->_head->as_CountedLoop();
156 Node *cl_exit = cl->loopexit();
157
158 // First clear the entries
159 for (uint i = 0; i < lpt()->_body.size(); i++) {
160 ignored_loop_nodes[i] = -1;
161 }
162
163 int max_vector = Matcher::max_vector_size(T_INT);
164
165 // Process the loop, some/all of the stack entries will not be in order, ergo
166 // need to preprocess the ignored initial state before we process the loop
167 for (uint i = 0; i < lpt()->_body.size(); i++) {
168 Node* n = lpt()->_body.at(i);
169 if (n == cl->incr() ||
170 n->is_reduction() ||
171 n->is_AddP() ||
172 n->is_Cmp() ||
173 n->is_IfTrue() ||
174 n->is_CountedLoop() ||
175 (n == cl_exit)) {
176 ignored_loop_nodes[i] = n->_idx;
177 continue;
178 }
179
180 if (n->is_If()) {
181 IfNode *iff = n->as_If();
182 if (iff->_fcnt != COUNT_UNKNOWN && iff->_prob != PROB_UNKNOWN) {
183 if (lpt()->is_loop_exit(iff)) {
184 ignored_loop_nodes[i] = n->_idx;
185 continue;
186 }
187 }
188 }
189
190 if (n->is_Phi() && (n->bottom_type() == Type::MEMORY)) {
191 Node* n_tail = n->in(LoopNode::LoopBackControl);
192 if (n_tail != n->in(LoopNode::EntryControl)) {
193 if (!n_tail->is_Mem()) {
194 is_slp = false;
195 break;
196 }
197 }
198 }
199
200 // This must happen after check of phi/if
201 if (n->is_Phi() || n->is_If()) {
202 ignored_loop_nodes[i] = n->_idx;
203 continue;
204 }
205
206 if (n->is_LoadStore() || n->is_MergeMem() ||
207 (n->is_Proj() && !n->as_Proj()->is_CFG())) {
208 is_slp = false;
209 break;
210 }
211
212 // Ignore nodes with non-primitive type.
213 BasicType bt;
214 if (n->is_Mem()) {
215 bt = n->as_Mem()->memory_type();
216 } else {
217 bt = n->bottom_type()->basic_type();
218 }
219 if (is_java_primitive(bt) == false) {
220 ignored_loop_nodes[i] = n->_idx;
221 continue;
222 }
223
224 if (n->is_Mem()) {
225 MemNode* current = n->as_Mem();
226 Node* adr = n->in(MemNode::Address);
227 Node* n_ctrl = _phase->get_ctrl(adr);
228
229 // save a queue of post process nodes
230 if (n_ctrl != NULL && lpt()->is_member(_phase->get_loop(n_ctrl))) {
231 // Process the memory expression
232 int stack_idx = 0;
233 bool have_side_effects = true;
234 if (adr->is_AddP() == false) {
235 nstack.push(adr, stack_idx++);
236 } else {
237 // Mark the components of the memory operation in nstack
238 SWPointer p1(current, this, &nstack, true);
239 have_side_effects = p1.node_stack()->is_nonempty();
240 }
241
242 // Process the pointer stack
243 while (have_side_effects) {
244 Node* pointer_node = nstack.node();
245 for (uint j = 0; j < lpt()->_body.size(); j++) {
246 Node* cur_node = lpt()->_body.at(j);
247 if (cur_node == pointer_node) {
248 ignored_loop_nodes[j] = cur_node->_idx;
249 break;
250 }
251 }
252 nstack.pop();
253 have_side_effects = nstack.is_nonempty();
254 }
255 }
256 }
257 }
258
259 if (is_slp) {
260 // Now we try to find the maximum supported consistent vector which the machine
261 // description can use
262 for (uint i = 0; i < lpt()->_body.size(); i++) {
263 if (ignored_loop_nodes[i] != -1) continue;
264
265 BasicType bt;
266 Node* n = lpt()->_body.at(i);
267 if (n->is_Mem()) {
268 bt = n->as_Mem()->memory_type();
269 } else {
270 bt = n->bottom_type()->basic_type();
271 }
272 if (is_java_primitive(bt) == false) continue;
273
274 int cur_max_vector = Matcher::max_vector_size(bt);
275
276 // If a max vector exists which is not larger than _local_loop_unroll_factor
277 // stop looking, we already have the max vector to map to.
278 if (cur_max_vector < local_loop_unroll_factor) {
279 is_slp = false;
280 if (TraceSuperWordLoopUnrollAnalysis) {
281 tty->print_cr("slp analysis fails: unroll limit greater than max vector\n");
282 }
283 break;
284 }
285
286 // Map the maximal common vector
287 if (VectorNode::implemented(n->Opcode(), cur_max_vector, bt)) {
288 if (cur_max_vector < max_vector) {
289 max_vector = cur_max_vector;
290 }
291 }
292 }
293 if (is_slp) {
294 local_loop_unroll_factor = max_vector;
295 cl->mark_passed_slp();
296 }
297 cl->mark_was_slp();
298 cl->set_slp_max_unroll(local_loop_unroll_factor);
299 }
300 }
301
302 //------------------------------SLP_extract---------------------------
303 // Extract the superword level parallelism
304 //
305 // 1) A reverse post-order of nodes in the block is constructed. By scanning
306 // this list from first to last, all definitions are visited before their uses.
307 //
308 // 2) A point-to-point dependence graph is constructed between memory references.
309 // This simplies the upcoming "independence" checker.
310 //
311 // 3) The maximum depth in the node graph from the beginning of the block
312 // to each node is computed. This is used to prune the graph search
313 // in the independence checker.
314 //
315 // 4) For integer types, the necessary bit width is propagated backwards
316 // from stores to allow packed operations on byte, char, and short
317 // integers. This reverses the promotion to type "int" that javac
318 // did for operations like: char c1,c2,c3; c1 = c2 + c3.
319 //
320 // 5) One of the memory references is picked to be an aligned vector reference.
321 // The pre-loop trip count is adjusted to align this reference in the
322 // unrolled body.
323 //
324 // 6) The initial set of pack pairs is seeded with memory references.
325 //
326 // 7) The set of pack pairs is extended by following use->def and def->use links.
327 //
328 // 8) The pairs are combined into vector sized packs.
329 //
330 // 9) Reorder the memory slices to co-locate members of the memory packs.
331 //
332 // 10) Generate ideal vector nodes for the final set of packs and where necessary,
333 // inserting scalar promotion, vector creation from multiple scalars, and
334 // extraction of scalar values from vectors.
335 //
336 void SuperWord::SLP_extract() {
337
338 #ifndef PRODUCT
339 if (_do_vector_loop && TraceSuperWord) {
340 tty->print("SuperWord::SLP_extract\n");
341 tty->print("input loop\n");
342 _lpt->dump_head();
343 _lpt->dump();
344 for (uint i = 0; i < _lpt->_body.size(); i++) {
345 _lpt->_body.at(i)->dump();
346 }
347 }
348 #endif
349 // Ready the block
350 if (!construct_bb()) {
351 return; // Exit if no interesting nodes or complex graph.
352 }
353 // build _dg, _disjoint_ptrs
354 dependence_graph();
355
356 // compute function depth(Node*)
357 compute_max_depth();
358
359 if (_do_vector_loop) {
360 if (mark_generations() != -1) {
361 hoist_loads_in_graph(); // this only rebuild the graph; all basic structs need rebuild explicitly
362
363 if (!construct_bb()) {
364 return; // Exit if no interesting nodes or complex graph.
365 }
366 dependence_graph();
367 compute_max_depth();
368 }
369
370 #ifndef PRODUCT
371 if (TraceSuperWord) {
372 tty->print_cr("\nSuperWord::_do_vector_loop: graph after hoist_loads_in_graph");
373 _lpt->dump_head();
374 for (int j = 0; j < _block.length(); j++) {
375 Node* n = _block.at(j);
376 int d = depth(n);
377 for (int i = 0; i < d; i++) tty->print("%s", " ");
378 tty->print("%d :", d);
379 n->dump();
380 }
381 }
382 #endif
383 }
384
385 compute_vector_element_type();
386
387 // Attempt vectorization
388
389 find_adjacent_refs();
390
391 extend_packlist();
392
393 if (_do_vector_loop) {
394 if (_packset.length() == 0) {
395 if (TraceSuperWord) {
396 tty->print_cr("\nSuperWord::_do_vector_loop DFA could not build packset, now trying to build anyway");
397 }
398 pack_parallel();
399 }
400 }
401
402 combine_packs();
403
404 construct_my_pack_map();
405
406 if (_do_vector_loop) {
407 merge_packs_to_cmovd();
408 }
409
410 filter_packs();
411
412 schedule();
413
414 output();
415 }
416
417 //------------------------------find_adjacent_refs---------------------------
418 // Find the adjacent memory references and create pack pairs for them.
419 // This is the initial set of packs that will then be extended by
420 // following use->def and def->use links. The align positions are
421 // assigned relative to the reference "align_to_ref"
422 void SuperWord::find_adjacent_refs() {
423 // Get list of memory operations
424 Node_List memops;
425 for (int i = 0; i < _block.length(); i++) {
426 Node* n = _block.at(i);
427 if (n->is_Mem() && !n->is_LoadStore() && in_bb(n) &&
428 is_java_primitive(n->as_Mem()->memory_type())) {
429 int align = memory_alignment(n->as_Mem(), 0);
430 if (align != bottom_align) {
431 memops.push(n);
432 }
433 }
434 }
435
436 Node_List align_to_refs;
437 int best_iv_adjustment = 0;
438 MemNode* best_align_to_mem_ref = NULL;
439
440 while (memops.size() != 0) {
441 // Find a memory reference to align to.
442 MemNode* mem_ref = find_align_to_ref(memops);
443 if (mem_ref == NULL) break;
444 align_to_refs.push(mem_ref);
445 int iv_adjustment = get_iv_adjustment(mem_ref);
446
447 if (best_align_to_mem_ref == NULL) {
448 // Set memory reference which is the best from all memory operations
449 // to be used for alignment. The pre-loop trip count is modified to align
450 // this reference to a vector-aligned address.
451 best_align_to_mem_ref = mem_ref;
452 best_iv_adjustment = iv_adjustment;
453 NOT_PRODUCT(find_adjacent_refs_trace_1(best_align_to_mem_ref, best_iv_adjustment);)
454 }
455
456 SWPointer align_to_ref_p(mem_ref, this, NULL, false);
457 // Set alignment relative to "align_to_ref" for all related memory operations.
458 for (int i = memops.size() - 1; i >= 0; i--) {
459 MemNode* s = memops.at(i)->as_Mem();
460 if (isomorphic(s, mem_ref) &&
461 (!_do_vector_loop || same_origin_idx(s, mem_ref))) {
462 SWPointer p2(s, this, NULL, false);
463 if (p2.comparable(align_to_ref_p)) {
464 int align = memory_alignment(s, iv_adjustment);
465 set_alignment(s, align);
466 }
467 }
468 }
469
470 // Create initial pack pairs of memory operations for which
471 // alignment is set and vectors will be aligned.
472 bool create_pack = true;
473 if (memory_alignment(mem_ref, best_iv_adjustment) == 0 || _do_vector_loop) {
474 if (!Matcher::misaligned_vectors_ok()) {
475 int vw = vector_width(mem_ref);
476 int vw_best = vector_width(best_align_to_mem_ref);
477 if (vw > vw_best) {
478 // Do not vectorize a memory access with more elements per vector
479 // if unaligned memory access is not allowed because number of
480 // iterations in pre-loop will be not enough to align it.
481 create_pack = false;
482 } else {
483 SWPointer p2(best_align_to_mem_ref, this, NULL, false);
484 if (align_to_ref_p.invar() != p2.invar()) {
485 // Do not vectorize memory accesses with different invariants
486 // if unaligned memory accesses are not allowed.
487 create_pack = false;
488 }
489 }
490 }
491 } else {
492 if (same_velt_type(mem_ref, best_align_to_mem_ref)) {
493 // Can't allow vectorization of unaligned memory accesses with the
494 // same type since it could be overlapped accesses to the same array.
495 create_pack = false;
496 } else {
497 // Allow independent (different type) unaligned memory operations
498 // if HW supports them.
499 if (!Matcher::misaligned_vectors_ok()) {
500 create_pack = false;
501 } else {
502 // Check if packs of the same memory type but
503 // with a different alignment were created before.
504 for (uint i = 0; i < align_to_refs.size(); i++) {
505 MemNode* mr = align_to_refs.at(i)->as_Mem();
506 if (same_velt_type(mr, mem_ref) &&
507 memory_alignment(mr, iv_adjustment) != 0)
508 create_pack = false;
509 }
510 }
511 }
512 }
513 if (create_pack) {
514 for (uint i = 0; i < memops.size(); i++) {
515 Node* s1 = memops.at(i);
516 int align = alignment(s1);
517 if (align == top_align) continue;
518 for (uint j = 0; j < memops.size(); j++) {
519 Node* s2 = memops.at(j);
520 if (alignment(s2) == top_align) continue;
521 if (s1 != s2 && are_adjacent_refs(s1, s2)) {
522 if (stmts_can_pack(s1, s2, align)) {
523 Node_List* pair = new Node_List();
524 pair->push(s1);
525 pair->push(s2);
526 if (!_do_vector_loop || same_origin_idx(s1, s2)) {
527 _packset.append(pair);
528 }
529 }
530 }
531 }
532 }
533 } else { // Don't create unaligned pack
534 // First, remove remaining memory ops of the same type from the list.
535 for (int i = memops.size() - 1; i >= 0; i--) {
536 MemNode* s = memops.at(i)->as_Mem();
537 if (same_velt_type(s, mem_ref)) {
538 memops.remove(i);
539 }
540 }
541
542 // Second, remove already constructed packs of the same type.
543 for (int i = _packset.length() - 1; i >= 0; i--) {
544 Node_List* p = _packset.at(i);
545 MemNode* s = p->at(0)->as_Mem();
546 if (same_velt_type(s, mem_ref)) {
547 remove_pack_at(i);
548 }
549 }
550
551 // If needed find the best memory reference for loop alignment again.
552 if (same_velt_type(mem_ref, best_align_to_mem_ref)) {
553 // Put memory ops from remaining packs back on memops list for
554 // the best alignment search.
555 uint orig_msize = memops.size();
556 for (int i = 0; i < _packset.length(); i++) {
557 Node_List* p = _packset.at(i);
558 MemNode* s = p->at(0)->as_Mem();
559 assert(!same_velt_type(s, mem_ref), "sanity");
560 memops.push(s);
561 }
562 best_align_to_mem_ref = find_align_to_ref(memops);
563 if (best_align_to_mem_ref == NULL) {
564 if (TraceSuperWord) {
565 tty->print_cr("SuperWord::find_adjacent_refs(): best_align_to_mem_ref == NULL");
566 }
567 break;
568 }
569 best_iv_adjustment = get_iv_adjustment(best_align_to_mem_ref);
570 NOT_PRODUCT(find_adjacent_refs_trace_1(best_align_to_mem_ref, best_iv_adjustment);)
571 // Restore list.
572 while (memops.size() > orig_msize)
573 (void)memops.pop();
574 }
575 } // unaligned memory accesses
576
577 // Remove used mem nodes.
578 for (int i = memops.size() - 1; i >= 0; i--) {
579 MemNode* m = memops.at(i)->as_Mem();
580 if (alignment(m) != top_align) {
581 memops.remove(i);
582 }
583 }
584
585 } // while (memops.size() != 0
586 set_align_to_ref(best_align_to_mem_ref);
587
588 if (TraceSuperWord) {
589 tty->print_cr("\nAfter find_adjacent_refs");
590 print_packset();
591 }
592 }
593
594 #ifndef PRODUCT
595 void SuperWord::find_adjacent_refs_trace_1(Node* best_align_to_mem_ref, int best_iv_adjustment) {
596 if (is_trace_adjacent()) {
597 tty->print("SuperWord::find_adjacent_refs best_align_to_mem_ref = %d, best_iv_adjustment = %d",
598 best_align_to_mem_ref->_idx, best_iv_adjustment);
599 best_align_to_mem_ref->dump();
600 }
601 }
602 #endif
603
604 //------------------------------find_align_to_ref---------------------------
605 // Find a memory reference to align the loop induction variable to.
606 // Looks first at stores then at loads, looking for a memory reference
607 // with the largest number of references similar to it.
608 MemNode* SuperWord::find_align_to_ref(Node_List &memops) {
609 GrowableArray<int> cmp_ct(arena(), memops.size(), memops.size(), 0);
610
611 // Count number of comparable memory ops
612 for (uint i = 0; i < memops.size(); i++) {
613 MemNode* s1 = memops.at(i)->as_Mem();
614 SWPointer p1(s1, this, NULL, false);
615 // Discard if pre loop can't align this reference
616 if (!ref_is_alignable(p1)) {
617 *cmp_ct.adr_at(i) = 0;
618 continue;
619 }
620 for (uint j = i+1; j < memops.size(); j++) {
621 MemNode* s2 = memops.at(j)->as_Mem();
622 if (isomorphic(s1, s2)) {
623 SWPointer p2(s2, this, NULL, false);
624 if (p1.comparable(p2)) {
625 (*cmp_ct.adr_at(i))++;
626 (*cmp_ct.adr_at(j))++;
627 }
628 }
629 }
630 }
631
632 // Find Store (or Load) with the greatest number of "comparable" references,
633 // biggest vector size, smallest data size and smallest iv offset.
634 int max_ct = 0;
635 int max_vw = 0;
636 int max_idx = -1;
637 int min_size = max_jint;
638 int min_iv_offset = max_jint;
639 for (uint j = 0; j < memops.size(); j++) {
640 MemNode* s = memops.at(j)->as_Mem();
641 if (s->is_Store()) {
642 int vw = vector_width_in_bytes(s);
643 assert(vw > 1, "sanity");
644 SWPointer p(s, this, NULL, false);
645 if (cmp_ct.at(j) > max_ct ||
646 cmp_ct.at(j) == max_ct &&
647 (vw > max_vw ||
648 vw == max_vw &&
649 (data_size(s) < min_size ||
650 data_size(s) == min_size &&
651 (p.offset_in_bytes() < min_iv_offset)))) {
652 max_ct = cmp_ct.at(j);
653 max_vw = vw;
654 max_idx = j;
655 min_size = data_size(s);
656 min_iv_offset = p.offset_in_bytes();
657 }
658 }
659 }
660 // If no stores, look at loads
661 if (max_ct == 0) {
662 for (uint j = 0; j < memops.size(); j++) {
663 MemNode* s = memops.at(j)->as_Mem();
664 if (s->is_Load()) {
665 int vw = vector_width_in_bytes(s);
666 assert(vw > 1, "sanity");
667 SWPointer p(s, this, NULL, false);
668 if (cmp_ct.at(j) > max_ct ||
669 cmp_ct.at(j) == max_ct &&
670 (vw > max_vw ||
671 vw == max_vw &&
672 (data_size(s) < min_size ||
673 data_size(s) == min_size &&
674 (p.offset_in_bytes() < min_iv_offset)))) {
675 max_ct = cmp_ct.at(j);
676 max_vw = vw;
677 max_idx = j;
678 min_size = data_size(s);
679 min_iv_offset = p.offset_in_bytes();
680 }
681 }
682 }
683 }
684
685 #ifdef ASSERT
686 if (TraceSuperWord && Verbose) {
687 tty->print_cr("\nVector memops after find_align_to_ref");
688 for (uint i = 0; i < memops.size(); i++) {
689 MemNode* s = memops.at(i)->as_Mem();
690 s->dump();
691 }
692 }
693 #endif
694
695 if (max_ct > 0) {
696 #ifdef ASSERT
697 if (TraceSuperWord) {
698 tty->print("\nVector align to node: ");
699 memops.at(max_idx)->as_Mem()->dump();
700 }
701 #endif
702 return memops.at(max_idx)->as_Mem();
703 }
704 return NULL;
705 }
706
707 //------------------------------ref_is_alignable---------------------------
708 // Can the preloop align the reference to position zero in the vector?
709 bool SuperWord::ref_is_alignable(SWPointer& p) {
710 if (!p.has_iv()) {
711 return true; // no induction variable
712 }
713 CountedLoopEndNode* pre_end = get_pre_loop_end(lp()->as_CountedLoop());
714 assert(pre_end != NULL, "we must have a correct pre-loop");
715 assert(pre_end->stride_is_con(), "pre loop stride is constant");
716 int preloop_stride = pre_end->stride_con();
717
718 int span = preloop_stride * p.scale_in_bytes();
719 int mem_size = p.memory_size();
720 int offset = p.offset_in_bytes();
721 // Stride one accesses are alignable if offset is aligned to memory operation size.
722 // Offset can be unaligned when UseUnalignedAccesses is used.
723 if (ABS(span) == mem_size && (ABS(offset) % mem_size) == 0) {
724 return true;
725 }
726 // If the initial offset from start of the object is computable,
727 // check if the pre-loop can align the final offset accordingly.
728 //
729 // In other words: Can we find an i such that the offset
730 // after i pre-loop iterations is aligned to vw?
731 // (init_offset + pre_loop) % vw == 0 (1)
732 // where
733 // pre_loop = i * span
734 // is the number of bytes added to the offset by i pre-loop iterations.
735 //
736 // For this to hold we need pre_loop to increase init_offset by
737 // pre_loop = vw - (init_offset % vw)
738 //
739 // This is only possible if pre_loop is divisible by span because each
740 // pre-loop iteration increases the initial offset by 'span' bytes:
741 // (vw - (init_offset % vw)) % span == 0
742 //
743 int vw = vector_width_in_bytes(p.mem());
744 assert(vw > 1, "sanity");
745 Node* init_nd = pre_end->init_trip();
746 if (init_nd->is_Con() && p.invar() == NULL) {
747 int init = init_nd->bottom_type()->is_int()->get_con();
748 int init_offset = init * p.scale_in_bytes() + offset;
749 assert(init_offset >= 0, "positive offset from object start");
750 if (vw % span == 0) {
751 // If vm is a multiple of span, we use formula (1).
752 if (span > 0) {
753 return (vw - (init_offset % vw)) % span == 0;
754 } else {
755 assert(span < 0, "nonzero stride * scale");
756 return (init_offset % vw) % -span == 0;
757 }
758 } else if (span % vw == 0) {
759 // If span is a multiple of vw, we can simplify formula (1) to:
760 // (init_offset + i * span) % vw == 0
761 // =>
762 // (init_offset % vw) + ((i * span) % vw) == 0
763 // =>
764 // init_offset % vw == 0
765 //
766 // Because we add a multiple of vw to the initial offset, the final
767 // offset is a multiple of vw if and only if init_offset is a multiple.
768 //
769 return (init_offset % vw) == 0;
770 }
771 }
772 return false;
773 }
774
775 //---------------------------get_iv_adjustment---------------------------
776 // Calculate loop's iv adjustment for this memory ops.
777 int SuperWord::get_iv_adjustment(MemNode* mem_ref) {
778 SWPointer align_to_ref_p(mem_ref, this, NULL, false);
779 int offset = align_to_ref_p.offset_in_bytes();
780 int scale = align_to_ref_p.scale_in_bytes();
781 int elt_size = align_to_ref_p.memory_size();
782 int vw = vector_width_in_bytes(mem_ref);
783 assert(vw > 1, "sanity");
784 int iv_adjustment;
785 if (scale != 0) {
786 int stride_sign = (scale * iv_stride()) > 0 ? 1 : -1;
787 // At least one iteration is executed in pre-loop by default. As result
788 // several iterations are needed to align memory operations in main-loop even
789 // if offset is 0.
790 int iv_adjustment_in_bytes = (stride_sign * vw - (offset % vw));
791 assert(((ABS(iv_adjustment_in_bytes) % elt_size) == 0),
792 "(%d) should be divisible by (%d)", iv_adjustment_in_bytes, elt_size);
793 iv_adjustment = iv_adjustment_in_bytes/elt_size;
794 } else {
795 // This memory op is not dependent on iv (scale == 0)
796 iv_adjustment = 0;
797 }
798
799 #ifndef PRODUCT
800 if (TraceSuperWord) {
801 tty->print("SuperWord::get_iv_adjustment: n = %d, noffset = %d iv_adjust = %d elt_size = %d scale = %d iv_stride = %d vect_size %d: ",
802 mem_ref->_idx, offset, iv_adjustment, elt_size, scale, iv_stride(), vw);
803 mem_ref->dump();
804 }
805 #endif
806 return iv_adjustment;
807 }
808
809 //---------------------------dependence_graph---------------------------
810 // Construct dependency graph.
811 // Add dependence edges to load/store nodes for memory dependence
812 // A.out()->DependNode.in(1) and DependNode.out()->B.prec(x)
813 void SuperWord::dependence_graph() {
814 // First, assign a dependence node to each memory node
815 for (int i = 0; i < _block.length(); i++ ) {
816 Node *n = _block.at(i);
817 if (n->is_Mem() || n->is_Phi() && n->bottom_type() == Type::MEMORY) {
818 _dg.make_node(n);
819 }
820 }
821
822 // For each memory slice, create the dependences
823 for (int i = 0; i < _mem_slice_head.length(); i++) {
824 Node* n = _mem_slice_head.at(i);
825 Node* n_tail = _mem_slice_tail.at(i);
826
827 // Get slice in predecessor order (last is first)
828 mem_slice_preds(n_tail, n, _nlist);
829
830 #ifndef PRODUCT
831 if(TraceSuperWord && Verbose) {
832 tty->print_cr("SuperWord::dependence_graph: built a new mem slice");
833 for (int j = _nlist.length() - 1; j >= 0 ; j--) {
834 _nlist.at(j)->dump();
835 }
836 }
837 #endif
838 // Make the slice dependent on the root
839 DepMem* slice = _dg.dep(n);
840 _dg.make_edge(_dg.root(), slice);
841
842 // Create a sink for the slice
843 DepMem* slice_sink = _dg.make_node(NULL);
844 _dg.make_edge(slice_sink, _dg.tail());
845
846 // Now visit each pair of memory ops, creating the edges
847 for (int j = _nlist.length() - 1; j >= 0 ; j--) {
848 Node* s1 = _nlist.at(j);
849
850 // If no dependency yet, use slice
851 if (_dg.dep(s1)->in_cnt() == 0) {
852 _dg.make_edge(slice, s1);
853 }
854 SWPointer p1(s1->as_Mem(), this, NULL, false);
855 bool sink_dependent = true;
856 for (int k = j - 1; k >= 0; k--) {
857 Node* s2 = _nlist.at(k);
858 if (s1->is_Load() && s2->is_Load())
859 continue;
860 SWPointer p2(s2->as_Mem(), this, NULL, false);
861
862 int cmp = p1.cmp(p2);
863 if (SuperWordRTDepCheck &&
864 p1.base() != p2.base() && p1.valid() && p2.valid()) {
865 // Create a runtime check to disambiguate
866 OrderedPair pp(p1.base(), p2.base());
867 _disjoint_ptrs.append_if_missing(pp);
868 } else if (!SWPointer::not_equal(cmp)) {
869 // Possibly same address
870 _dg.make_edge(s1, s2);
871 sink_dependent = false;
872 }
873 }
874 if (sink_dependent) {
875 _dg.make_edge(s1, slice_sink);
876 }
877 }
878
879 if (TraceSuperWord) {
880 tty->print_cr("\nDependence graph for slice: %d", n->_idx);
881 for (int q = 0; q < _nlist.length(); q++) {
882 _dg.print(_nlist.at(q));
883 }
884 tty->cr();
885 }
886
887 _nlist.clear();
888 }
889
890 if (TraceSuperWord) {
891 tty->print_cr("\ndisjoint_ptrs: %s", _disjoint_ptrs.length() > 0 ? "" : "NONE");
892 for (int r = 0; r < _disjoint_ptrs.length(); r++) {
893 _disjoint_ptrs.at(r).print();
894 tty->cr();
895 }
896 tty->cr();
897 }
898
899 }
900
901 //---------------------------mem_slice_preds---------------------------
902 // Return a memory slice (node list) in predecessor order starting at "start"
903 void SuperWord::mem_slice_preds(Node* start, Node* stop, GrowableArray<Node*> &preds) {
904 assert(preds.length() == 0, "start empty");
905 Node* n = start;
906 Node* prev = NULL;
907 while (true) {
908 NOT_PRODUCT( if(is_trace_mem_slice()) tty->print_cr("SuperWord::mem_slice_preds: n %d", n->_idx);)
909 assert(in_bb(n), "must be in block");
910 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
911 Node* out = n->fast_out(i);
912 if (out->is_Load()) {
913 if (in_bb(out)) {
914 preds.push(out);
915 if (TraceSuperWord && Verbose) {
916 tty->print_cr("SuperWord::mem_slice_preds: added pred(%d)", out->_idx);
917 }
918 }
919 } else {
920 // FIXME
921 if (out->is_MergeMem() && !in_bb(out)) {
922 // Either unrolling is causing a memory edge not to disappear,
923 // or need to run igvn.optimize() again before SLP
924 } else if (out->is_Phi() && out->bottom_type() == Type::MEMORY && !in_bb(out)) {
925 // Ditto. Not sure what else to check further.
926 } else if (out->Opcode() == Op_StoreCM && out->in(MemNode::OopStore) == n) {
927 // StoreCM has an input edge used as a precedence edge.
928 // Maybe an issue when oop stores are vectorized.
929 } else {
930 assert(out == prev || prev == NULL, "no branches off of store slice");
931 }
932 }//else
933 }//for
934 if (n == stop) break;
935 preds.push(n);
936 if (TraceSuperWord && Verbose) {
937 tty->print_cr("SuperWord::mem_slice_preds: added pred(%d)", n->_idx);
938 }
939 prev = n;
940 assert(n->is_Mem(), "unexpected node %s", n->Name());
941 n = n->in(MemNode::Memory);
942 }
943 }
944
945 //------------------------------stmts_can_pack---------------------------
946 // Can s1 and s2 be in a pack with s1 immediately preceding s2 and
947 // s1 aligned at "align"
948 bool SuperWord::stmts_can_pack(Node* s1, Node* s2, int align) {
949
950 // Do not use superword for non-primitives
951 BasicType bt1 = velt_basic_type(s1);
952 BasicType bt2 = velt_basic_type(s2);
953 if(!is_java_primitive(bt1) || !is_java_primitive(bt2))
954 return false;
955 if (Matcher::max_vector_size(bt1) < 2) {
956 return false; // No vectors for this type
957 }
958
959 if (isomorphic(s1, s2)) {
960 if (independent(s1, s2) || reduction(s1, s2)) {
961 if (!exists_at(s1, 0) && !exists_at(s2, 1)) {
962 if (!s1->is_Mem() || are_adjacent_refs(s1, s2)) {
963 int s1_align = alignment(s1);
964 int s2_align = alignment(s2);
965 if (s1_align == top_align || s1_align == align) {
966 if (s2_align == top_align || s2_align == align + data_size(s1)) {
967 return true;
968 }
969 }
970 }
971 }
972 }
973 }
974 return false;
975 }
976
977 //------------------------------exists_at---------------------------
978 // Does s exist in a pack at position pos?
979 bool SuperWord::exists_at(Node* s, uint pos) {
980 for (int i = 0; i < _packset.length(); i++) {
981 Node_List* p = _packset.at(i);
982 if (p->at(pos) == s) {
983 return true;
984 }
985 }
986 return false;
987 }
988
989 //------------------------------are_adjacent_refs---------------------------
990 // Is s1 immediately before s2 in memory?
991 bool SuperWord::are_adjacent_refs(Node* s1, Node* s2) {
992 if (!s1->is_Mem() || !s2->is_Mem()) return false;
993 if (!in_bb(s1) || !in_bb(s2)) return false;
994
995 // Do not use superword for non-primitives
996 if (!is_java_primitive(s1->as_Mem()->memory_type()) ||
997 !is_java_primitive(s2->as_Mem()->memory_type())) {
998 return false;
999 }
1000
1001 // FIXME - co_locate_pack fails on Stores in different mem-slices, so
1002 // only pack memops that are in the same alias set until that's fixed.
1003 if (_phase->C->get_alias_index(s1->as_Mem()->adr_type()) !=
1004 _phase->C->get_alias_index(s2->as_Mem()->adr_type()))
1005 return false;
1006 SWPointer p1(s1->as_Mem(), this, NULL, false);
1007 SWPointer p2(s2->as_Mem(), this, NULL, false);
1008 if (p1.base() != p2.base() || !p1.comparable(p2)) return false;
1009 int diff = p2.offset_in_bytes() - p1.offset_in_bytes();
1010 return diff == data_size(s1);
1011 }
1012
1013 //------------------------------isomorphic---------------------------
1014 // Are s1 and s2 similar?
1015 bool SuperWord::isomorphic(Node* s1, Node* s2) {
1016 if (s1->Opcode() != s2->Opcode()) return false;
1017 if (s1->req() != s2->req()) return false;
1018 if (s1->in(0) != s2->in(0)) return false;
1019 if (!same_velt_type(s1, s2)) return false;
1020 return true;
1021 }
1022
1023 //------------------------------independent---------------------------
1024 // Is there no data path from s1 to s2 or s2 to s1?
1025 bool SuperWord::independent(Node* s1, Node* s2) {
1026 // assert(s1->Opcode() == s2->Opcode(), "check isomorphic first");
1027 int d1 = depth(s1);
1028 int d2 = depth(s2);
1029 if (d1 == d2) return s1 != s2;
1030 Node* deep = d1 > d2 ? s1 : s2;
1031 Node* shallow = d1 > d2 ? s2 : s1;
1032
1033 visited_clear();
1034
1035 return independent_path(shallow, deep);
1036 }
1037
1038 //------------------------------reduction---------------------------
1039 // Is there a data path between s1 and s2 and the nodes reductions?
1040 bool SuperWord::reduction(Node* s1, Node* s2) {
1041 bool retValue = false;
1042 int d1 = depth(s1);
1043 int d2 = depth(s2);
1044 if (d1 + 1 == d2) {
1045 if (s1->is_reduction() && s2->is_reduction()) {
1046 // This is an ordered set, so s1 should define s2
1047 for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
1048 Node* t1 = s1->fast_out(i);
1049 if (t1 == s2) {
1050 // both nodes are reductions and connected
1051 retValue = true;
1052 }
1053 }
1054 }
1055 }
1056
1057 return retValue;
1058 }
1059
1060 //------------------------------independent_path------------------------------
1061 // Helper for independent
1062 bool SuperWord::independent_path(Node* shallow, Node* deep, uint dp) {
1063 if (dp >= 1000) return false; // stop deep recursion
1064 visited_set(deep);
1065 int shal_depth = depth(shallow);
1066 assert(shal_depth <= depth(deep), "must be");
1067 for (DepPreds preds(deep, _dg); !preds.done(); preds.next()) {
1068 Node* pred = preds.current();
1069 if (in_bb(pred) && !visited_test(pred)) {
1070 if (shallow == pred) {
1071 return false;
1072 }
1073 if (shal_depth < depth(pred) && !independent_path(shallow, pred, dp+1)) {
1074 return false;
1075 }
1076 }
1077 }
1078 return true;
1079 }
1080
1081 //------------------------------set_alignment---------------------------
1082 void SuperWord::set_alignment(Node* s1, Node* s2, int align) {
1083 set_alignment(s1, align);
1084 if (align == top_align || align == bottom_align) {
1085 set_alignment(s2, align);
1086 } else {
1087 set_alignment(s2, align + data_size(s1));
1088 }
1089 }
1090
1091 //------------------------------data_size---------------------------
1092 int SuperWord::data_size(Node* s) {
1093 Node* use = NULL; //test if the node is a candidate for CMoveVD optimization, then return the size of CMov
1094 if (_do_vector_loop) {
1095 use = _cmovev_kit.is_Bool_candidate(s);
1096 if (use != NULL) {
1097 return data_size(use);
1098 }
1099 use = _cmovev_kit.is_CmpD_candidate(s);
1100 if (use != NULL) {
1101 return data_size(use);
1102 }
1103 }
1104 int bsize = type2aelembytes(velt_basic_type(s));
1105 assert(bsize != 0, "valid size");
1106 return bsize;
1107 }
1108
1109 //------------------------------extend_packlist---------------------------
1110 // Extend packset by following use->def and def->use links from pack members.
1111 void SuperWord::extend_packlist() {
1112 bool changed;
1113 do {
1114 packset_sort(_packset.length());
1115 changed = false;
1116 for (int i = 0; i < _packset.length(); i++) {
1117 Node_List* p = _packset.at(i);
1118 changed |= follow_use_defs(p);
1119 changed |= follow_def_uses(p);
1120 }
1121 } while (changed);
1122
1123 if (_race_possible) {
1124 for (int i = 0; i < _packset.length(); i++) {
1125 Node_List* p = _packset.at(i);
1126 order_def_uses(p);
1127 }
1128 }
1129
1130 if (TraceSuperWord) {
1131 tty->print_cr("\nAfter extend_packlist");
1132 print_packset();
1133 }
1134 }
1135
1136 //------------------------------follow_use_defs---------------------------
1137 // Extend the packset by visiting operand definitions of nodes in pack p
1138 bool SuperWord::follow_use_defs(Node_List* p) {
1139 assert(p->size() == 2, "just checking");
1140 Node* s1 = p->at(0);
1141 Node* s2 = p->at(1);
1142 assert(s1->req() == s2->req(), "just checking");
1143 assert(alignment(s1) + data_size(s1) == alignment(s2), "just checking");
1144
1145 if (s1->is_Load()) return false;
1146
1147 int align = alignment(s1);
1148 NOT_PRODUCT(if(is_trace_alignment()) tty->print_cr("SuperWord::follow_use_defs: s1 %d, align %d", s1->_idx, align);)
1149 bool changed = false;
1150 int start = s1->is_Store() ? MemNode::ValueIn : 1;
1151 int end = s1->is_Store() ? MemNode::ValueIn+1 : s1->req();
1152 for (int j = start; j < end; j++) {
1153 Node* t1 = s1->in(j);
1154 Node* t2 = s2->in(j);
1155 if (!in_bb(t1) || !in_bb(t2))
1156 continue;
1157 if (stmts_can_pack(t1, t2, align)) {
1158 if (est_savings(t1, t2) >= 0) {
1159 Node_List* pair = new Node_List();
1160 pair->push(t1);
1161 pair->push(t2);
1162 _packset.append(pair);
1163 NOT_PRODUCT(if(is_trace_alignment()) tty->print_cr("SuperWord::follow_use_defs: set_alignment(%d, %d, %d)", t1->_idx, t2->_idx, align);)
1164 set_alignment(t1, t2, align);
1165 changed = true;
1166 }
1167 }
1168 }
1169 return changed;
1170 }
1171
1172 //------------------------------follow_def_uses---------------------------
1173 // Extend the packset by visiting uses of nodes in pack p
1174 bool SuperWord::follow_def_uses(Node_List* p) {
1175 bool changed = false;
1176 Node* s1 = p->at(0);
1177 Node* s2 = p->at(1);
1178 assert(p->size() == 2, "just checking");
1179 assert(s1->req() == s2->req(), "just checking");
1180 assert(alignment(s1) + data_size(s1) == alignment(s2), "just checking");
1181
1182 if (s1->is_Store()) return false;
1183
1184 int align = alignment(s1);
1185 NOT_PRODUCT(if(is_trace_alignment()) tty->print_cr("SuperWord::follow_def_uses: s1 %d, align %d", s1->_idx, align);)
1186 int savings = -1;
1187 int num_s1_uses = 0;
1188 Node* u1 = NULL;
1189 Node* u2 = NULL;
1190 for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
1191 Node* t1 = s1->fast_out(i);
1192 num_s1_uses++;
1193 if (!in_bb(t1)) continue;
1194 for (DUIterator_Fast jmax, j = s2->fast_outs(jmax); j < jmax; j++) {
1195 Node* t2 = s2->fast_out(j);
1196 if (!in_bb(t2)) continue;
1197 if (!opnd_positions_match(s1, t1, s2, t2))
1198 continue;
1199 if (stmts_can_pack(t1, t2, align)) {
1200 int my_savings = est_savings(t1, t2);
1201 if (my_savings > savings) {
1202 savings = my_savings;
1203 u1 = t1;
1204 u2 = t2;
1205 }
1206 }
1207 }
1208 }
1209 if (num_s1_uses > 1) {
1210 _race_possible = true;
1211 }
1212 if (savings >= 0) {
1213 Node_List* pair = new Node_List();
1214 pair->push(u1);
1215 pair->push(u2);
1216 _packset.append(pair);
1217 NOT_PRODUCT(if(is_trace_alignment()) tty->print_cr("SuperWord::follow_def_uses: set_alignment(%d, %d, %d)", u1->_idx, u2->_idx, align);)
1218 set_alignment(u1, u2, align);
1219 changed = true;
1220 }
1221 return changed;
1222 }
1223
1224 //------------------------------order_def_uses---------------------------
1225 // For extended packsets, ordinally arrange uses packset by major component
1226 void SuperWord::order_def_uses(Node_List* p) {
1227 Node* s1 = p->at(0);
1228
1229 if (s1->is_Store()) return;
1230
1231 // reductions are always managed beforehand
1232 if (s1->is_reduction()) return;
1233
1234 for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
1235 Node* t1 = s1->fast_out(i);
1236
1237 // Only allow operand swap on commuting operations
1238 if (!t1->is_Add() && !t1->is_Mul()) {
1239 break;
1240 }
1241
1242 // Now find t1's packset
1243 Node_List* p2 = NULL;
1244 for (int j = 0; j < _packset.length(); j++) {
1245 p2 = _packset.at(j);
1246 Node* first = p2->at(0);
1247 if (t1 == first) {
1248 break;
1249 }
1250 p2 = NULL;
1251 }
1252 // Arrange all sub components by the major component
1253 if (p2 != NULL) {
1254 for (uint j = 1; j < p->size(); j++) {
1255 Node* d1 = p->at(j);
1256 Node* u1 = p2->at(j);
1257 opnd_positions_match(s1, t1, d1, u1);
1258 }
1259 }
1260 }
1261 }
1262
1263 //---------------------------opnd_positions_match-------------------------
1264 // Is the use of d1 in u1 at the same operand position as d2 in u2?
1265 bool SuperWord::opnd_positions_match(Node* d1, Node* u1, Node* d2, Node* u2) {
1266 // check reductions to see if they are marshalled to represent the reduction
1267 // operator in a specified opnd
1268 if (u1->is_reduction() && u2->is_reduction()) {
1269 // ensure reductions have phis and reduction definitions feeding the 1st operand
1270 Node* first = u1->in(2);
1271 if (first->is_Phi() || first->is_reduction()) {
1272 u1->swap_edges(1, 2);
1273 }
1274 // ensure reductions have phis and reduction definitions feeding the 1st operand
1275 first = u2->in(2);
1276 if (first->is_Phi() || first->is_reduction()) {
1277 u2->swap_edges(1, 2);
1278 }
1279 return true;
1280 }
1281
1282 uint ct = u1->req();
1283 if (ct != u2->req()) return false;
1284 uint i1 = 0;
1285 uint i2 = 0;
1286 do {
1287 for (i1++; i1 < ct; i1++) if (u1->in(i1) == d1) break;
1288 for (i2++; i2 < ct; i2++) if (u2->in(i2) == d2) break;
1289 if (i1 != i2) {
1290 if ((i1 == (3-i2)) && (u2->is_Add() || u2->is_Mul())) {
1291 // Further analysis relies on operands position matching.
1292 u2->swap_edges(i1, i2);
1293 } else {
1294 return false;
1295 }
1296 }
1297 } while (i1 < ct);
1298 return true;
1299 }
1300
1301 //------------------------------est_savings---------------------------
1302 // Estimate the savings from executing s1 and s2 as a pack
1303 int SuperWord::est_savings(Node* s1, Node* s2) {
1304 int save_in = 2 - 1; // 2 operations per instruction in packed form
1305
1306 // inputs
1307 for (uint i = 1; i < s1->req(); i++) {
1308 Node* x1 = s1->in(i);
1309 Node* x2 = s2->in(i);
1310 if (x1 != x2) {
1311 if (are_adjacent_refs(x1, x2)) {
1312 save_in += adjacent_profit(x1, x2);
1313 } else if (!in_packset(x1, x2)) {
1314 save_in -= pack_cost(2);
1315 } else {
1316 save_in += unpack_cost(2);
1317 }
1318 }
1319 }
1320
1321 // uses of result
1322 uint ct = 0;
1323 int save_use = 0;
1324 for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
1325 Node* s1_use = s1->fast_out(i);
1326 for (int j = 0; j < _packset.length(); j++) {
1327 Node_List* p = _packset.at(j);
1328 if (p->at(0) == s1_use) {
1329 for (DUIterator_Fast kmax, k = s2->fast_outs(kmax); k < kmax; k++) {
1330 Node* s2_use = s2->fast_out(k);
1331 if (p->at(p->size()-1) == s2_use) {
1332 ct++;
1333 if (are_adjacent_refs(s1_use, s2_use)) {
1334 save_use += adjacent_profit(s1_use, s2_use);
1335 }
1336 }
1337 }
1338 }
1339 }
1340 }
1341
1342 if (ct < s1->outcnt()) save_use += unpack_cost(1);
1343 if (ct < s2->outcnt()) save_use += unpack_cost(1);
1344
1345 return MAX2(save_in, save_use);
1346 }
1347
1348 //------------------------------costs---------------------------
1349 int SuperWord::adjacent_profit(Node* s1, Node* s2) { return 2; }
1350 int SuperWord::pack_cost(int ct) { return ct; }
1351 int SuperWord::unpack_cost(int ct) { return ct; }
1352
1353 //------------------------------combine_packs---------------------------
1354 // Combine packs A and B with A.last == B.first into A.first..,A.last,B.second,..B.last
1355 void SuperWord::combine_packs() {
1356 bool changed = true;
1357 // Combine packs regardless max vector size.
1358 while (changed) {
1359 changed = false;
1360 for (int i = 0; i < _packset.length(); i++) {
1361 Node_List* p1 = _packset.at(i);
1362 if (p1 == NULL) continue;
1363 // Because of sorting we can start at i + 1
1364 for (int j = i + 1; j < _packset.length(); j++) {
1365 Node_List* p2 = _packset.at(j);
1366 if (p2 == NULL) continue;
1367 if (i == j) continue;
1368 if (p1->at(p1->size()-1) == p2->at(0)) {
1369 for (uint k = 1; k < p2->size(); k++) {
1370 p1->push(p2->at(k));
1371 }
1372 _packset.at_put(j, NULL);
1373 changed = true;
1374 }
1375 }
1376 }
1377 }
1378
1379 // Split packs which have size greater then max vector size.
1380 for (int i = 0; i < _packset.length(); i++) {
1381 Node_List* p1 = _packset.at(i);
1382 if (p1 != NULL) {
1383 BasicType bt = velt_basic_type(p1->at(0));
1384 uint max_vlen = Matcher::max_vector_size(bt); // Max elements in vector
1385 assert(is_power_of_2(max_vlen), "sanity");
1386 uint psize = p1->size();
1387 if (!is_power_of_2(psize)) {
1388 // Skip pack which can't be vector.
1389 // case1: for(...) { a[i] = i; } elements values are different (i+x)
1390 // case2: for(...) { a[i] = b[i+1]; } can't align both, load and store
1391 _packset.at_put(i, NULL);
1392 continue;
1393 }
1394 if (psize > max_vlen) {
1395 Node_List* pack = new Node_List();
1396 for (uint j = 0; j < psize; j++) {
1397 pack->push(p1->at(j));
1398 if (pack->size() >= max_vlen) {
1399 assert(is_power_of_2(pack->size()), "sanity");
1400 _packset.append(pack);
1401 pack = new Node_List();
1402 }
1403 }
1404 _packset.at_put(i, NULL);
1405 }
1406 }
1407 }
1408
1409 // Compress list.
1410 for (int i = _packset.length() - 1; i >= 0; i--) {
1411 Node_List* p1 = _packset.at(i);
1412 if (p1 == NULL) {
1413 _packset.remove_at(i);
1414 }
1415 }
1416
1417 if (TraceSuperWord) {
1418 tty->print_cr("\nAfter combine_packs");
1419 print_packset();
1420 }
1421 }
1422
1423 //-----------------------------construct_my_pack_map--------------------------
1424 // Construct the map from nodes to packs. Only valid after the
1425 // point where a node is only in one pack (after combine_packs).
1426 void SuperWord::construct_my_pack_map() {
1427 Node_List* rslt = NULL;
1428 for (int i = 0; i < _packset.length(); i++) {
1429 Node_List* p = _packset.at(i);
1430 for (uint j = 0; j < p->size(); j++) {
1431 Node* s = p->at(j);
1432 assert(my_pack(s) == NULL, "only in one pack");
1433 set_my_pack(s, p);
1434 }
1435 }
1436 }
1437
1438 //------------------------------filter_packs---------------------------
1439 // Remove packs that are not implemented or not profitable.
1440 void SuperWord::filter_packs() {
1441 // Remove packs that are not implemented
1442 for (int i = _packset.length() - 1; i >= 0; i--) {
1443 Node_List* pk = _packset.at(i);
1444 bool impl = implemented(pk);
1445 if (!impl) {
1446 #ifndef PRODUCT
1447 if (TraceSuperWord && Verbose) {
1448 tty->print_cr("Unimplemented");
1449 pk->at(0)->dump();
1450 }
1451 #endif
1452 remove_pack_at(i);
1453 }
1454 Node *n = pk->at(0);
1455 if (n->is_reduction()) {
1456 _num_reductions++;
1457 } else {
1458 _num_work_vecs++;
1459 }
1460 }
1461
1462 // Remove packs that are not profitable
1463 bool changed;
1464 do {
1465 changed = false;
1466 for (int i = _packset.length() - 1; i >= 0; i--) {
1467 Node_List* pk = _packset.at(i);
1468 bool prof = profitable(pk);
1469 if (!prof) {
1470 #ifndef PRODUCT
1471 if (TraceSuperWord && Verbose) {
1472 tty->print_cr("Unprofitable");
1473 pk->at(0)->dump();
1474 }
1475 #endif
1476 remove_pack_at(i);
1477 changed = true;
1478 }
1479 }
1480 } while (changed);
1481
1482 #ifndef PRODUCT
1483 if (TraceSuperWord) {
1484 tty->print_cr("\nAfter filter_packs");
1485 print_packset();
1486 tty->cr();
1487 }
1488 #endif
1489 }
1490
1491 //------------------------------merge_packs_to_cmovd---------------------------
1492 // Merge CMoveD into new vector-nodes
1493 // We want to catch this pattern and subsume CmpD and Bool into CMoveD
1494 //
1495 // SubD ConD
1496 // / | /
1497 // / | / /
1498 // / | / /
1499 // / | / /
1500 // / / /
1501 // / / | /
1502 // v / | /
1503 // CmpD | /
1504 // | | /
1505 // v | /
1506 // Bool | /
1507 // \ | /
1508 // \ | /
1509 // \ | /
1510 // \ | /
1511 // \ v /
1512 // CMoveD
1513 //
1514
1515 void SuperWord::merge_packs_to_cmovd() {
1516 for (int i = _packset.length() - 1; i >= 0; i--) {
1517 _cmovev_kit.make_cmovevd_pack(_packset.at(i));
1518 }
1519 #ifndef PRODUCT
1520 if (TraceSuperWord) {
1521 tty->print_cr("\nSuperWord::merge_packs_to_cmovd(): After merge");
1522 print_packset();
1523 tty->cr();
1524 }
1525 #endif
1526 }
1527
1528 Node* CMoveKit::is_Bool_candidate(Node* def) const {
1529 Node* use = NULL;
1530 if (!def->is_Bool() || def->in(0) != NULL || def->outcnt() != 1) {
1531 return NULL;
1532 }
1533 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
1534 use = def->fast_out(j);
1535 if (!_sw->same_generation(def, use) || !use->is_CMove()) {
1536 return NULL;
1537 }
1538 }
1539 return use;
1540 }
1541
1542 Node* CMoveKit::is_CmpD_candidate(Node* def) const {
1543 Node* use = NULL;
1544 if (!def->is_Cmp() || def->in(0) != NULL || def->outcnt() != 1) {
1545 return NULL;
1546 }
1547 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
1548 use = def->fast_out(j);
1549 if (!_sw->same_generation(def, use) || (use = is_Bool_candidate(use)) == NULL || !_sw->same_generation(def, use)) {
1550 return NULL;
1551 }
1552 }
1553 return use;
1554 }
1555
1556 Node_List* CMoveKit::make_cmovevd_pack(Node_List* cmovd_pk) {
1557 Node *cmovd = cmovd_pk->at(0);
1558 if (!cmovd->is_CMove()) {
1559 return NULL;
1560 }
1561 if (pack(cmovd) != NULL) { // already in the cmov pack
1562 return NULL;
1563 }
1564 if (cmovd->in(0) != NULL) {
1565 NOT_PRODUCT(if(_sw->is_trace_cmov()) {tty->print("CMoveKit::make_cmovevd_pack: CMoveD %d has control flow, escaping...", cmovd->_idx); cmovd->dump();})
1566 return NULL;
1567 }
1568
1569 Node* bol = cmovd->as_CMove()->in(CMoveNode::Condition);
1570 if (!bol->is_Bool()
1571 || bol->outcnt() != 1
1572 || !_sw->same_generation(bol, cmovd)
1573 || bol->in(0) != NULL // BoolNode has control flow!!
1574 || _sw->my_pack(bol) == NULL) {
1575 NOT_PRODUCT(if(_sw->is_trace_cmov()) {tty->print("CMoveKit::make_cmovevd_pack: Bool %d does not fit CMoveD %d for building vector, escaping...", bol->_idx, cmovd->_idx); bol->dump();})
1576 return NULL;
1577 }
1578 Node_List* bool_pk = _sw->my_pack(bol);
1579 if (bool_pk->size() != cmovd_pk->size() ) {
1580 return NULL;
1581 }
1582
1583 Node* cmpd = bol->in(1);
1584 if (!cmpd->is_Cmp()
1585 || cmpd->outcnt() != 1
1586 || !_sw->same_generation(cmpd, cmovd)
1587 || cmpd->in(0) != NULL // CmpDNode has control flow!!
1588 || _sw->my_pack(cmpd) == NULL) {
1589 NOT_PRODUCT(if(_sw->is_trace_cmov()) {tty->print("CMoveKit::make_cmovevd_pack: CmpD %d does not fit CMoveD %d for building vector, escaping...", cmpd->_idx, cmovd->_idx); cmpd->dump();})
1590 return NULL;
1591 }
1592 Node_List* cmpd_pk = _sw->my_pack(cmpd);
1593 if (cmpd_pk->size() != cmovd_pk->size() ) {
1594 return NULL;
1595 }
1596
1597 if (!test_cmpd_pack(cmpd_pk, cmovd_pk)) {
1598 NOT_PRODUCT(if(_sw->is_trace_cmov()) {tty->print("CMoveKit::make_cmovevd_pack: cmpd pack for CmpD %d failed vectorization test", cmpd->_idx); cmpd->dump();})
1599 return NULL;
1600 }
1601
1602 Node_List* new_cmpd_pk = new Node_List();
1603 uint sz = cmovd_pk->size() - 1;
1604 for (uint i = 0; i <= sz; ++i) {
1605 Node* cmov = cmovd_pk->at(i);
1606 Node* bol = bool_pk->at(i);
1607 Node* cmp = cmpd_pk->at(i);
1608
1609 new_cmpd_pk->insert(i, cmov);
1610
1611 map(cmov, new_cmpd_pk);
1612 map(bol, new_cmpd_pk);
1613 map(cmp, new_cmpd_pk);
1614
1615 _sw->set_my_pack(cmov, new_cmpd_pk); // and keep old packs for cmp and bool
1616 }
1617 _sw->_packset.remove(cmovd_pk);
1618 _sw->_packset.remove(bool_pk);
1619 _sw->_packset.remove(cmpd_pk);
1620 _sw->_packset.append(new_cmpd_pk);
1621 NOT_PRODUCT(if(_sw->is_trace_cmov()) {tty->print_cr("CMoveKit::make_cmovevd_pack: added syntactic CMoveD pack"); _sw->print_pack(new_cmpd_pk);})
1622 return new_cmpd_pk;
1623 }
1624
1625 bool CMoveKit::test_cmpd_pack(Node_List* cmpd_pk, Node_List* cmovd_pk) {
1626 Node* cmpd0 = cmpd_pk->at(0);
1627 assert(cmpd0->is_Cmp(), "CMoveKit::test_cmpd_pack: should be CmpDNode");
1628 assert(cmovd_pk->at(0)->is_CMove(), "CMoveKit::test_cmpd_pack: should be CMoveD");
1629 assert(cmpd_pk->size() == cmovd_pk->size(), "CMoveKit::test_cmpd_pack: should be same size");
1630 Node* in1 = cmpd0->in(1);
1631 Node* in2 = cmpd0->in(2);
1632 Node_List* in1_pk = _sw->my_pack(in1);
1633 Node_List* in2_pk = _sw->my_pack(in2);
1634
1635 if (in1_pk != NULL && in1_pk->size() != cmpd_pk->size()
1636 || in2_pk != NULL && in2_pk->size() != cmpd_pk->size() ) {
1637 return false;
1638 }
1639
1640 // test if "all" in1 are in the same pack or the same node
1641 if (in1_pk == NULL) {
1642 for (uint j = 1; j < cmpd_pk->size(); j++) {
1643 if (cmpd_pk->at(j)->in(1) != in1) {
1644 return false;
1645 }
1646 }//for: in1_pk is not pack but all CmpD nodes in the pack have the same in(1)
1647 }
1648 // test if "all" in2 are in the same pack or the same node
1649 if (in2_pk == NULL) {
1650 for (uint j = 1; j < cmpd_pk->size(); j++) {
1651 if (cmpd_pk->at(j)->in(2) != in2) {
1652 return false;
1653 }
1654 }//for: in2_pk is not pack but all CmpD nodes in the pack have the same in(2)
1655 }
1656 //now check if cmpd_pk may be subsumed in vector built for cmovd_pk
1657 int cmovd_ind1, cmovd_ind2;
1658 if (cmpd_pk->at(0)->in(1) == cmovd_pk->at(0)->as_CMove()->in(CMoveNode::IfFalse)
1659 && cmpd_pk->at(0)->in(2) == cmovd_pk->at(0)->as_CMove()->in(CMoveNode::IfTrue)) {
1660 cmovd_ind1 = CMoveNode::IfFalse;
1661 cmovd_ind2 = CMoveNode::IfTrue;
1662 } else if (cmpd_pk->at(0)->in(2) == cmovd_pk->at(0)->as_CMove()->in(CMoveNode::IfFalse)
1663 && cmpd_pk->at(0)->in(1) == cmovd_pk->at(0)->as_CMove()->in(CMoveNode::IfTrue)) {
1664 cmovd_ind2 = CMoveNode::IfFalse;
1665 cmovd_ind1 = CMoveNode::IfTrue;
1666 }
1667 else {
1668 return false;
1669 }
1670
1671 for (uint j = 1; j < cmpd_pk->size(); j++) {
1672 if (cmpd_pk->at(j)->in(1) != cmovd_pk->at(j)->as_CMove()->in(cmovd_ind1)
1673 || cmpd_pk->at(j)->in(2) != cmovd_pk->at(j)->as_CMove()->in(cmovd_ind2)) {
1674 return false;
1675 }//if
1676 }
1677 NOT_PRODUCT(if(_sw->is_trace_cmov()) { tty->print("CMoveKit::test_cmpd_pack: cmpd pack for 1st CmpD %d is OK for vectorization: ", cmpd0->_idx); cmpd0->dump(); })
1678 return true;
1679 }
1680
1681 //------------------------------implemented---------------------------
1682 // Can code be generated for pack p?
1683 bool SuperWord::implemented(Node_List* p) {
1684 bool retValue = false;
1685 Node* p0 = p->at(0);
1686 if (p0 != NULL) {
1687 int opc = p0->Opcode();
1688 uint size = p->size();
1689 if (p0->is_reduction()) {
1690 const Type *arith_type = p0->bottom_type();
1691 // Length 2 reductions of INT/LONG do not offer performance benefits
1692 if (((arith_type->basic_type() == T_INT) || (arith_type->basic_type() == T_LONG)) && (size == 2)) {
1693 retValue = false;
1694 } else {
1695 retValue = ReductionNode::implemented(opc, size, arith_type->basic_type());
1696 }
1697 } else {
1698 retValue = VectorNode::implemented(opc, size, velt_basic_type(p0));
1699 }
1700 if (!retValue) {
1701 if (is_cmov_pack(p)) {
1702 NOT_PRODUCT(if(is_trace_cmov()) {tty->print_cr("SWPointer::implemented: found cmpd pack"); print_pack(p);})
1703 return true;
1704 }
1705 }
1706 }
1707 return retValue;
1708 }
1709
1710 bool SuperWord::is_cmov_pack(Node_List* p) {
1711 return _cmovev_kit.pack(p->at(0)) != NULL;
1712 }
1713 //------------------------------same_inputs--------------------------
1714 // For pack p, are all idx operands the same?
1715 bool SuperWord::same_inputs(Node_List* p, int idx) {
1716 Node* p0 = p->at(0);
1717 uint vlen = p->size();
1718 Node* p0_def = p0->in(idx);
1719 for (uint i = 1; i < vlen; i++) {
1720 Node* pi = p->at(i);
1721 Node* pi_def = pi->in(idx);
1722 if (p0_def != pi_def) {
1723 return false;
1724 }
1725 }
1726 return true;
1727 }
1728
1729 //------------------------------profitable---------------------------
1730 // For pack p, are all operands and all uses (with in the block) vector?
1731 bool SuperWord::profitable(Node_List* p) {
1732 Node* p0 = p->at(0);
1733 uint start, end;
1734 VectorNode::vector_operands(p0, &start, &end);
1735
1736 // Return false if some inputs are not vectors or vectors with different
1737 // size or alignment.
1738 // Also, for now, return false if not scalar promotion case when inputs are
1739 // the same. Later, implement PackNode and allow differing, non-vector inputs
1740 // (maybe just the ones from outside the block.)
1741 for (uint i = start; i < end; i++) {
1742 if (!is_vector_use(p0, i)) {
1743 return false;
1744 }
1745 }
1746 // Check if reductions are connected
1747 if (p0->is_reduction()) {
1748 Node* second_in = p0->in(2);
1749 Node_List* second_pk = my_pack(second_in);
1750 if ((second_pk == NULL) || (_num_work_vecs == _num_reductions)) {
1751 // Remove reduction flag if no parent pack or if not enough work
1752 // to cover reduction expansion overhead
1753 p0->remove_flag(Node::Flag_is_reduction);
1754 return false;
1755 } else if (second_pk->size() != p->size()) {
1756 return false;
1757 }
1758 }
1759 if (VectorNode::is_shift(p0)) {
1760 // For now, return false if shift count is vector or not scalar promotion
1761 // case (different shift counts) because it is not supported yet.
1762 Node* cnt = p0->in(2);
1763 Node_List* cnt_pk = my_pack(cnt);
1764 if (cnt_pk != NULL)
1765 return false;
1766 if (!same_inputs(p, 2))
1767 return false;
1768 }
1769 if (!p0->is_Store()) {
1770 // For now, return false if not all uses are vector.
1771 // Later, implement ExtractNode and allow non-vector uses (maybe
1772 // just the ones outside the block.)
1773 for (uint i = 0; i < p->size(); i++) {
1774 Node* def = p->at(i);
1775 if (is_cmov_pack_internal_node(p, def)) {
1776 continue;
1777 }
1778 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
1779 Node* use = def->fast_out(j);
1780 for (uint k = 0; k < use->req(); k++) {
1781 Node* n = use->in(k);
1782 if (def == n) {
1783 // reductions can be loop carried dependences
1784 if (def->is_reduction() && use->is_Phi())
1785 continue;
1786 if (!is_vector_use(use, k)) {
1787 return false;
1788 }
1789 }
1790 }
1791 }
1792 }
1793 }
1794 return true;
1795 }
1796
1797 //------------------------------schedule---------------------------
1798 // Adjust the memory graph for the packed operations
1799 void SuperWord::schedule() {
1800
1801 // Co-locate in the memory graph the members of each memory pack
1802 for (int i = 0; i < _packset.length(); i++) {
1803 co_locate_pack(_packset.at(i));
1804 }
1805 }
1806
1807 //-------------------------------remove_and_insert-------------------
1808 // Remove "current" from its current position in the memory graph and insert
1809 // it after the appropriate insertion point (lip or uip).
1810 void SuperWord::remove_and_insert(MemNode *current, MemNode *prev, MemNode *lip,
1811 Node *uip, Unique_Node_List &sched_before) {
1812 Node* my_mem = current->in(MemNode::Memory);
1813 bool sched_up = sched_before.member(current);
1814
1815 // remove current_store from its current position in the memmory graph
1816 for (DUIterator i = current->outs(); current->has_out(i); i++) {
1817 Node* use = current->out(i);
1818 if (use->is_Mem()) {
1819 assert(use->in(MemNode::Memory) == current, "must be");
1820 if (use == prev) { // connect prev to my_mem
1821 _igvn.replace_input_of(use, MemNode::Memory, my_mem);
1822 --i; //deleted this edge; rescan position
1823 } else if (sched_before.member(use)) {
1824 if (!sched_up) { // Will be moved together with current
1825 _igvn.replace_input_of(use, MemNode::Memory, uip);
1826 --i; //deleted this edge; rescan position
1827 }
1828 } else {
1829 if (sched_up) { // Will be moved together with current
1830 _igvn.replace_input_of(use, MemNode::Memory, lip);
1831 --i; //deleted this edge; rescan position
1832 }
1833 }
1834 }
1835 }
1836
1837 Node *insert_pt = sched_up ? uip : lip;
1838
1839 // all uses of insert_pt's memory state should use current's instead
1840 for (DUIterator i = insert_pt->outs(); insert_pt->has_out(i); i++) {
1841 Node* use = insert_pt->out(i);
1842 if (use->is_Mem()) {
1843 assert(use->in(MemNode::Memory) == insert_pt, "must be");
1844 _igvn.replace_input_of(use, MemNode::Memory, current);
1845 --i; //deleted this edge; rescan position
1846 } else if (!sched_up && use->is_Phi() && use->bottom_type() == Type::MEMORY) {
1847 uint pos; //lip (lower insert point) must be the last one in the memory slice
1848 for (pos=1; pos < use->req(); pos++) {
1849 if (use->in(pos) == insert_pt) break;
1850 }
1851 _igvn.replace_input_of(use, pos, current);
1852 --i;
1853 }
1854 }
1855
1856 //connect current to insert_pt
1857 _igvn.replace_input_of(current, MemNode::Memory, insert_pt);
1858 }
1859
1860 //------------------------------co_locate_pack----------------------------------
1861 // To schedule a store pack, we need to move any sandwiched memory ops either before
1862 // or after the pack, based upon dependence information:
1863 // (1) If any store in the pack depends on the sandwiched memory op, the
1864 // sandwiched memory op must be scheduled BEFORE the pack;
1865 // (2) If a sandwiched memory op depends on any store in the pack, the
1866 // sandwiched memory op must be scheduled AFTER the pack;
1867 // (3) If a sandwiched memory op (say, memA) depends on another sandwiched
1868 // memory op (say memB), memB must be scheduled before memA. So, if memA is
1869 // scheduled before the pack, memB must also be scheduled before the pack;
1870 // (4) If there is no dependence restriction for a sandwiched memory op, we simply
1871 // schedule this store AFTER the pack
1872 // (5) We know there is no dependence cycle, so there in no other case;
1873 // (6) Finally, all memory ops in another single pack should be moved in the same direction.
1874 //
1875 // To schedule a load pack, we use the memory state of either the first or the last load in
1876 // the pack, based on the dependence constraint.
1877 void SuperWord::co_locate_pack(Node_List* pk) {
1878 if (pk->at(0)->is_Store()) {
1879 MemNode* first = executed_first(pk)->as_Mem();
1880 MemNode* last = executed_last(pk)->as_Mem();
1881 Unique_Node_List schedule_before_pack;
1882 Unique_Node_List memops;
1883
1884 MemNode* current = last->in(MemNode::Memory)->as_Mem();
1885 MemNode* previous = last;
1886 while (true) {
1887 assert(in_bb(current), "stay in block");
1888 memops.push(previous);
1889 for (DUIterator i = current->outs(); current->has_out(i); i++) {
1890 Node* use = current->out(i);
1891 if (use->is_Mem() && use != previous)
1892 memops.push(use);
1893 }
1894 if (current == first) break;
1895 previous = current;
1896 current = current->in(MemNode::Memory)->as_Mem();
1897 }
1898
1899 // determine which memory operations should be scheduled before the pack
1900 for (uint i = 1; i < memops.size(); i++) {
1901 Node *s1 = memops.at(i);
1902 if (!in_pack(s1, pk) && !schedule_before_pack.member(s1)) {
1903 for (uint j = 0; j< i; j++) {
1904 Node *s2 = memops.at(j);
1905 if (!independent(s1, s2)) {
1906 if (in_pack(s2, pk) || schedule_before_pack.member(s2)) {
1907 schedule_before_pack.push(s1); // s1 must be scheduled before
1908 Node_List* mem_pk = my_pack(s1);
1909 if (mem_pk != NULL) {
1910 for (uint ii = 0; ii < mem_pk->size(); ii++) {
1911 Node* s = mem_pk->at(ii); // follow partner
1912 if (memops.member(s) && !schedule_before_pack.member(s))
1913 schedule_before_pack.push(s);
1914 }
1915 }
1916 break;
1917 }
1918 }
1919 }
1920 }
1921 }
1922
1923 Node* upper_insert_pt = first->in(MemNode::Memory);
1924 // Following code moves loads connected to upper_insert_pt below aliased stores.
1925 // Collect such loads here and reconnect them back to upper_insert_pt later.
1926 memops.clear();
1927 for (DUIterator i = upper_insert_pt->outs(); upper_insert_pt->has_out(i); i++) {
1928 Node* use = upper_insert_pt->out(i);
1929 if (use->is_Mem() && !use->is_Store()) {
1930 memops.push(use);
1931 }
1932 }
1933
1934 MemNode* lower_insert_pt = last;
1935 previous = last; //previous store in pk
1936 current = last->in(MemNode::Memory)->as_Mem();
1937
1938 // start scheduling from "last" to "first"
1939 while (true) {
1940 assert(in_bb(current), "stay in block");
1941 assert(in_pack(previous, pk), "previous stays in pack");
1942 Node* my_mem = current->in(MemNode::Memory);
1943
1944 if (in_pack(current, pk)) {
1945 // Forward users of my memory state (except "previous) to my input memory state
1946 for (DUIterator i = current->outs(); current->has_out(i); i++) {
1947 Node* use = current->out(i);
1948 if (use->is_Mem() && use != previous) {
1949 assert(use->in(MemNode::Memory) == current, "must be");
1950 if (schedule_before_pack.member(use)) {
1951 _igvn.replace_input_of(use, MemNode::Memory, upper_insert_pt);
1952 } else {
1953 _igvn.replace_input_of(use, MemNode::Memory, lower_insert_pt);
1954 }
1955 --i; // deleted this edge; rescan position
1956 }
1957 }
1958 previous = current;
1959 } else { // !in_pack(current, pk) ==> a sandwiched store
1960 remove_and_insert(current, previous, lower_insert_pt, upper_insert_pt, schedule_before_pack);
1961 }
1962
1963 if (current == first) break;
1964 current = my_mem->as_Mem();
1965 } // end while
1966
1967 // Reconnect loads back to upper_insert_pt.
1968 for (uint i = 0; i < memops.size(); i++) {
1969 Node *ld = memops.at(i);
1970 if (ld->in(MemNode::Memory) != upper_insert_pt) {
1971 _igvn.replace_input_of(ld, MemNode::Memory, upper_insert_pt);
1972 }
1973 }
1974 } else if (pk->at(0)->is_Load()) { //load
1975 // all loads in the pack should have the same memory state. By default,
1976 // we use the memory state of the last load. However, if any load could
1977 // not be moved down due to the dependence constraint, we use the memory
1978 // state of the first load.
1979 Node* last_mem = executed_last(pk)->in(MemNode::Memory);
1980 Node* first_mem = executed_first(pk)->in(MemNode::Memory);
1981 bool schedule_last = true;
1982 for (uint i = 0; i < pk->size(); i++) {
1983 Node* ld = pk->at(i);
1984 for (Node* current = last_mem; current != ld->in(MemNode::Memory);
1985 current=current->in(MemNode::Memory)) {
1986 assert(current != first_mem, "corrupted memory graph");
1987 if(current->is_Mem() && !independent(current, ld)){
1988 schedule_last = false; // a later store depends on this load
1989 break;
1990 }
1991 }
1992 }
1993
1994 Node* mem_input = schedule_last ? last_mem : first_mem;
1995 _igvn.hash_delete(mem_input);
1996 // Give each load the same memory state
1997 for (uint i = 0; i < pk->size(); i++) {
1998 LoadNode* ld = pk->at(i)->as_Load();
1999 _igvn.replace_input_of(ld, MemNode::Memory, mem_input);
2000 }
2001 }
2002 }
2003
2004 #ifndef PRODUCT
2005 void SuperWord::print_loop(bool whole) {
2006 Node_Stack stack(_arena, _phase->C->unique() >> 2);
2007 Node_List rpo_list;
2008 VectorSet visited(_arena);
2009 visited.set(lpt()->_head->_idx);
2010 _phase->rpo(lpt()->_head, stack, visited, rpo_list);
2011 _phase->dump(lpt(), rpo_list.size(), rpo_list );
2012 if(whole) {
2013 tty->print_cr("\n Whole loop tree");
2014 _phase->dump();
2015 tty->print_cr(" End of whole loop tree\n");
2016 }
2017 }
2018 #endif
2019
2020 //------------------------------output---------------------------
2021 // Convert packs into vector node operations
2022 void SuperWord::output() {
2023 if (_packset.length() == 0) return;
2024
2025 #ifndef PRODUCT
2026 if (TraceLoopOpts) {
2027 tty->print("SuperWord::output ");
2028 lpt()->dump_head();
2029 }
2030 #endif
2031
2032 // MUST ENSURE main loop's initial value is properly aligned:
2033 // (iv_initial_value + min_iv_offset) % vector_width_in_bytes() == 0
2034
2035 align_initial_loop_index(align_to_ref());
2036
2037 // Insert extract (unpack) operations for scalar uses
2038 for (int i = 0; i < _packset.length(); i++) {
2039 insert_extracts(_packset.at(i));
2040 }
2041
2042 Compile* C = _phase->C;
2043 CountedLoopNode *cl = lpt()->_head->as_CountedLoop();
2044 uint max_vlen_in_bytes = 0;
2045 uint max_vlen = 0;
2046
2047 NOT_PRODUCT(if(is_trace_loop_reverse()) {tty->print_cr("SWPointer::output: print loop before create_reserve_version_of_loop"); print_loop(true);})
2048
2049 CountedLoopReserveKit make_reversable(_phase, _lpt, do_reserve_copy());
2050
2051 NOT_PRODUCT(if(is_trace_loop_reverse()) {tty->print_cr("SWPointer::output: print loop after create_reserve_version_of_loop"); print_loop(true);})
2052
2053 if (do_reserve_copy() && !make_reversable.has_reserved()) {
2054 NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: loop was not reserved correctly, exiting SuperWord");})
2055 return;
2056 }
2057
2058 for (int i = 0; i < _block.length(); i++) {
2059 Node* n = _block.at(i);
2060 Node_List* p = my_pack(n);
2061 if (p && n == executed_last(p)) {
2062 uint vlen = p->size();
2063 uint vlen_in_bytes = 0;
2064 Node* vn = NULL;
2065 Node* low_adr = p->at(0);
2066 Node* first = executed_first(p);
2067 NOT_PRODUCT(if(is_trace_cmov()) {tty->print_cr("SWPointer::output: %d executed first, %d executed last in pack", first->_idx, n->_idx); print_pack(p);})
2068 int opc = n->Opcode();
2069 if (n->is_Load()) {
2070 Node* ctl = n->in(MemNode::Control);
2071 Node* mem = first->in(MemNode::Memory);
2072 SWPointer p1(n->as_Mem(), this, NULL, false);
2073 // Identify the memory dependency for the new loadVector node by
2074 // walking up through memory chain.
2075 // This is done to give flexibility to the new loadVector node so that
2076 // it can move above independent storeVector nodes.
2077 while (mem->is_StoreVector()) {
2078 SWPointer p2(mem->as_Mem(), this, NULL, false);
2079 int cmp = p1.cmp(p2);
2080 if (SWPointer::not_equal(cmp) || !SWPointer::comparable(cmp)) {
2081 mem = mem->in(MemNode::Memory);
2082 } else {
2083 break; // dependent memory
2084 }
2085 }
2086 Node* adr = low_adr->in(MemNode::Address);
2087 const TypePtr* atyp = n->adr_type();
2088 vn = LoadVectorNode::make(opc, ctl, mem, adr, atyp, vlen, velt_basic_type(n), control_dependency(p));
2089 vlen_in_bytes = vn->as_LoadVector()->memory_size();
2090 } else if (n->is_Store()) {
2091 // Promote value to be stored to vector
2092 Node* val = vector_opd(p, MemNode::ValueIn);
2093 if (val == NULL) {
2094 if (do_reserve_copy()) {
2095 NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: val should not be NULL, exiting SuperWord");})
2096 return; //and reverse to backup IG
2097 }
2098 ShouldNotReachHere();
2099 }
2100
2101 Node* ctl = n->in(MemNode::Control);
2102 Node* mem = first->in(MemNode::Memory);
2103 Node* adr = low_adr->in(MemNode::Address);
2104 const TypePtr* atyp = n->adr_type();
2105 vn = StoreVectorNode::make(opc, ctl, mem, adr, atyp, val, vlen);
2106 vlen_in_bytes = vn->as_StoreVector()->memory_size();
2107 } else if (n->req() == 3 && !is_cmov_pack(p)) {
2108 // Promote operands to vector
2109 Node* in1 = NULL;
2110 bool node_isa_reduction = n->is_reduction();
2111 if (node_isa_reduction) {
2112 // the input to the first reduction operation is retained
2113 in1 = low_adr->in(1);
2114 } else {
2115 in1 = vector_opd(p, 1);
2116 if (in1 == NULL) {
2117 if (do_reserve_copy()) {
2118 NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: in1 should not be NULL, exiting SuperWord");})
2119 return; //and reverse to backup IG
2120 }
2121 ShouldNotReachHere();
2122 }
2123 }
2124 Node* in2 = vector_opd(p, 2);
2125 if (in2 == NULL) {
2126 if (do_reserve_copy()) {
2127 NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: in2 should not be NULL, exiting SuperWord");})
2128 return; //and reverse to backup IG
2129 }
2130 ShouldNotReachHere();
2131 }
2132 if (VectorNode::is_invariant_vector(in1) && (node_isa_reduction == false) && (n->is_Add() || n->is_Mul())) {
2133 // Move invariant vector input into second position to avoid register spilling.
2134 Node* tmp = in1;
2135 in1 = in2;
2136 in2 = tmp;
2137 }
2138 if (node_isa_reduction) {
2139 const Type *arith_type = n->bottom_type();
2140 vn = ReductionNode::make(opc, NULL, in1, in2, arith_type->basic_type());
2141 if (in2->is_Load()) {
2142 vlen_in_bytes = in2->as_LoadVector()->memory_size();
2143 } else {
2144 vlen_in_bytes = in2->as_Vector()->length_in_bytes();
2145 }
2146 } else {
2147 vn = VectorNode::make(opc, in1, in2, vlen, velt_basic_type(n));
2148 vlen_in_bytes = vn->as_Vector()->length_in_bytes();
2149 }
2150 } else if (opc == Op_SqrtD || opc == Op_AbsF || opc == Op_AbsD || opc == Op_NegF || opc == Op_NegD) {
2151 // Promote operand to vector (Sqrt/Abs/Neg are 2 address instructions)
2152 Node* in = vector_opd(p, 1);
2153 vn = VectorNode::make(opc, in, NULL, vlen, velt_basic_type(n));
2154 vlen_in_bytes = vn->as_Vector()->length_in_bytes();
2155 } else if (is_cmov_pack(p)) {
2156 if (!n->is_CMove()) {
2157 continue;
2158 }
2159 // place here CMoveVDNode
2160 NOT_PRODUCT(if(is_trace_cmov()) {tty->print_cr("SWPointer::output: print before CMove vectorization"); print_loop(false);})
2161 Node* bol = n->in(CMoveNode::Condition);
2162 if (!bol->is_Bool() && bol->Opcode() == Op_ExtractI && bol->req() > 1 ) {
2163 NOT_PRODUCT(if(is_trace_cmov()) {tty->print_cr("SWPointer::output: %d is not Bool node, trying its in(1) node %d", bol->_idx, bol->in(1)->_idx); bol->dump(); bol->in(1)->dump();})
2164 bol = bol->in(1); //may be ExtractNode
2165 }
2166
2167 assert(bol->is_Bool(), "should be BoolNode - too late to bail out!");
2168 if (!bol->is_Bool()) {
2169 if (do_reserve_copy()) {
2170 NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: expected %d bool node, exiting SuperWord", bol->_idx); bol->dump();})
2171 return; //and reverse to backup IG
2172 }
2173 ShouldNotReachHere();
2174 }
2175
2176 int cond = (int)bol->as_Bool()->_test._test;
2177 Node* in_cc = _igvn.intcon(cond);
2178 NOT_PRODUCT(if(is_trace_cmov()) {tty->print("SWPointer::output: created intcon in_cc node %d", in_cc->_idx); in_cc->dump();})
2179 Node* cc = bol->clone();
2180 cc->set_req(1, in_cc);
2181 NOT_PRODUCT(if(is_trace_cmov()) {tty->print("SWPointer::output: created bool cc node %d", cc->_idx); cc->dump();})
2182
2183 Node* src1 = vector_opd(p, 2); //2=CMoveNode::IfFalse
2184 if (src1 == NULL) {
2185 if (do_reserve_copy()) {
2186 NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: src1 should not be NULL, exiting SuperWord");})
2187 return; //and reverse to backup IG
2188 }
2189 ShouldNotReachHere();
2190 }
2191 Node* src2 = vector_opd(p, 3); //3=CMoveNode::IfTrue
2192 if (src2 == NULL) {
2193 if (do_reserve_copy()) {
2194 NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: src2 should not be NULL, exiting SuperWord");})
2195 return; //and reverse to backup IG
2196 }
2197 ShouldNotReachHere();
2198 }
2199 BasicType bt = velt_basic_type(n);
2200 const TypeVect* vt = TypeVect::make(bt, vlen);
2201 vn = new CMoveVDNode(cc, src1, src2, vt);
2202 NOT_PRODUCT(if(is_trace_cmov()) {tty->print("SWPointer::output: created new CMove node %d: ", vn->_idx); vn->dump();})
2203 } else {
2204 if (do_reserve_copy()) {
2205 NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: ShouldNotReachHere, exiting SuperWord");})
2206 return; //and reverse to backup IG
2207 }
2208 ShouldNotReachHere();
2209 }
2210
2211 assert(vn != NULL, "sanity");
2212 if (vn == NULL) {
2213 if (do_reserve_copy()){
2214 NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: got NULL node, cannot proceed, exiting SuperWord");})
2215 return; //and reverse to backup IG
2216 }
2217 ShouldNotReachHere();
2218 }
2219
2220 _igvn.register_new_node_with_optimizer(vn);
2221 _phase->set_ctrl(vn, _phase->get_ctrl(p->at(0)));
2222 for (uint j = 0; j < p->size(); j++) {
2223 Node* pm = p->at(j);
2224 _igvn.replace_node(pm, vn);
2225 }
2226 _igvn._worklist.push(vn);
2227
2228 if (vlen_in_bytes > max_vlen_in_bytes) {
2229 max_vlen = vlen;
2230 max_vlen_in_bytes = vlen_in_bytes;
2231 }
2232 #ifdef ASSERT
2233 if (TraceNewVectors) {
2234 tty->print("new Vector node: ");
2235 vn->dump();
2236 }
2237 #endif
2238 }
2239 }//for (int i = 0; i < _block.length(); i++)
2240
2241 C->set_max_vector_size(max_vlen_in_bytes);
2242
2243 if (SuperWordLoopUnrollAnalysis) {
2244 if (cl->has_passed_slp()) {
2245 uint slp_max_unroll_factor = cl->slp_max_unroll();
2246 if (slp_max_unroll_factor == max_vlen) {
2247 if (TraceSuperWordLoopUnrollAnalysis) {
2248 tty->print_cr("vector loop(unroll=%d, len=%d)\n", max_vlen, max_vlen_in_bytes*BitsPerByte);
2249 }
2250 // For atomic unrolled loops which are vector mapped, instigate more unrolling.
2251 cl->set_notpassed_slp();
2252 // if vector resources are limited, do not allow additional unrolling
2253 if (FLOATPRESSURE > 8) {
2254 C->set_major_progress();
2255 }
2256 cl->mark_do_unroll_only();
2257 if (do_reserve_copy()) {
2258 cl->mark_loop_vectorized();
2259 }
2260 }
2261 }
2262 }
2263
2264 if (do_reserve_copy()) {
2265 make_reversable.use_new();
2266 }
2267 NOT_PRODUCT(if(is_trace_loop_reverse()) {tty->print_cr("\n Final loop after SuperWord"); print_loop(true);})
2268 return;
2269 }
2270
2271 //------------------------------vector_opd---------------------------
2272 // Create a vector operand for the nodes in pack p for operand: in(opd_idx)
2273 Node* SuperWord::vector_opd(Node_List* p, int opd_idx) {
2274 Node* p0 = p->at(0);
2275 uint vlen = p->size();
2276 Node* opd = p0->in(opd_idx);
2277
2278 if (same_inputs(p, opd_idx)) {
2279 if (opd->is_Vector() || opd->is_LoadVector()) {
2280 assert(((opd_idx != 2) || !VectorNode::is_shift(p0)), "shift's count can't be vector");
2281 if (opd_idx == 2 && VectorNode::is_shift(p0)) {
2282 NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("shift's count can't be vector");})
2283 return NULL;
2284 }
2285 return opd; // input is matching vector
2286 }
2287 if ((opd_idx == 2) && VectorNode::is_shift(p0)) {
2288 Compile* C = _phase->C;
2289 Node* cnt = opd;
2290 // Vector instructions do not mask shift count, do it here.
2291 juint mask = (p0->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
2292 const TypeInt* t = opd->find_int_type();
2293 if (t != NULL && t->is_con()) {
2294 juint shift = t->get_con();
2295 if (shift > mask) { // Unsigned cmp
2296 cnt = ConNode::make(TypeInt::make(shift & mask));
2297 }
2298 } else {
2299 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
2300 cnt = ConNode::make(TypeInt::make(mask));
2301 _igvn.register_new_node_with_optimizer(cnt);
2302 cnt = new AndINode(opd, cnt);
2303 _igvn.register_new_node_with_optimizer(cnt);
2304 _phase->set_ctrl(cnt, _phase->get_ctrl(opd));
2305 }
2306 assert(opd->bottom_type()->isa_int(), "int type only");
2307 if (!opd->bottom_type()->isa_int()) {
2308 NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("Should be int type only");})
2309 return NULL;
2310 }
2311 // Move non constant shift count into vector register.
2312 cnt = VectorNode::shift_count(p0, cnt, vlen, velt_basic_type(p0));
2313 }
2314 if (cnt != opd) {
2315 _igvn.register_new_node_with_optimizer(cnt);
2316 _phase->set_ctrl(cnt, _phase->get_ctrl(opd));
2317 }
2318 return cnt;
2319 }
2320 assert(!opd->is_StoreVector(), "such vector is not expected here");
2321 if (opd->is_StoreVector()) {
2322 NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("StoreVector is not expected here");})
2323 return NULL;
2324 }
2325 // Convert scalar input to vector with the same number of elements as
2326 // p0's vector. Use p0's type because size of operand's container in
2327 // vector should match p0's size regardless operand's size.
2328 const Type* p0_t = velt_type(p0);
2329 VectorNode* vn = VectorNode::scalar2vector(opd, vlen, p0_t);
2330
2331 _igvn.register_new_node_with_optimizer(vn);
2332 _phase->set_ctrl(vn, _phase->get_ctrl(opd));
2333 #ifdef ASSERT
2334 if (TraceNewVectors) {
2335 tty->print("new Vector node: ");
2336 vn->dump();
2337 }
2338 #endif
2339 return vn;
2340 }
2341
2342 // Insert pack operation
2343 BasicType bt = velt_basic_type(p0);
2344 PackNode* pk = PackNode::make(opd, vlen, bt);
2345 DEBUG_ONLY( const BasicType opd_bt = opd->bottom_type()->basic_type(); )
2346
2347 for (uint i = 1; i < vlen; i++) {
2348 Node* pi = p->at(i);
2349 Node* in = pi->in(opd_idx);
2350 assert(my_pack(in) == NULL, "Should already have been unpacked");
2351 if (my_pack(in) != NULL) {
2352 NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("Should already have been unpacked");})
2353 return NULL;
2354 }
2355 assert(opd_bt == in->bottom_type()->basic_type(), "all same type");
2356 pk->add_opd(in);
2357 }
2358 _igvn.register_new_node_with_optimizer(pk);
2359 _phase->set_ctrl(pk, _phase->get_ctrl(opd));
2360 #ifdef ASSERT
2361 if (TraceNewVectors) {
2362 tty->print("new Vector node: ");
2363 pk->dump();
2364 }
2365 #endif
2366 return pk;
2367 }
2368
2369 //------------------------------insert_extracts---------------------------
2370 // If a use of pack p is not a vector use, then replace the
2371 // use with an extract operation.
2372 void SuperWord::insert_extracts(Node_List* p) {
2373 if (p->at(0)->is_Store()) return;
2374 assert(_n_idx_list.is_empty(), "empty (node,index) list");
2375
2376 // Inspect each use of each pack member. For each use that is
2377 // not a vector use, replace the use with an extract operation.
2378
2379 for (uint i = 0; i < p->size(); i++) {
2380 Node* def = p->at(i);
2381 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
2382 Node* use = def->fast_out(j);
2383 for (uint k = 0; k < use->req(); k++) {
2384 Node* n = use->in(k);
2385 if (def == n) {
2386 Node_List* u_pk = my_pack(use);
2387 if ((u_pk == NULL || !is_cmov_pack(u_pk) || use->is_CMove()) && !is_vector_use(use, k)) {
2388 _n_idx_list.push(use, k);
2389 }
2390 }
2391 }
2392 }
2393 }
2394
2395 while (_n_idx_list.is_nonempty()) {
2396 Node* use = _n_idx_list.node();
2397 int idx = _n_idx_list.index();
2398 _n_idx_list.pop();
2399 Node* def = use->in(idx);
2400
2401 if (def->is_reduction()) continue;
2402
2403 // Insert extract operation
2404 _igvn.hash_delete(def);
2405 int def_pos = alignment(def) / data_size(def);
2406
2407 Node* ex = ExtractNode::make(def, def_pos, velt_basic_type(def));
2408 _igvn.register_new_node_with_optimizer(ex);
2409 _phase->set_ctrl(ex, _phase->get_ctrl(def));
2410 _igvn.replace_input_of(use, idx, ex);
2411 _igvn._worklist.push(def);
2412
2413 bb_insert_after(ex, bb_idx(def));
2414 set_velt_type(ex, velt_type(def));
2415 }
2416 }
2417
2418 //------------------------------is_vector_use---------------------------
2419 // Is use->in(u_idx) a vector use?
2420 bool SuperWord::is_vector_use(Node* use, int u_idx) {
2421 Node_List* u_pk = my_pack(use);
2422 if (u_pk == NULL) return false;
2423 if (use->is_reduction()) return true;
2424 Node* def = use->in(u_idx);
2425 Node_List* d_pk = my_pack(def);
2426 if (d_pk == NULL) {
2427 // check for scalar promotion
2428 Node* n = u_pk->at(0)->in(u_idx);
2429 for (uint i = 1; i < u_pk->size(); i++) {
2430 if (u_pk->at(i)->in(u_idx) != n) return false;
2431 }
2432 return true;
2433 }
2434 if (u_pk->size() != d_pk->size())
2435 return false;
2436 for (uint i = 0; i < u_pk->size(); i++) {
2437 Node* ui = u_pk->at(i);
2438 Node* di = d_pk->at(i);
2439 if (ui->in(u_idx) != di || alignment(ui) != alignment(di))
2440 return false;
2441 }
2442 return true;
2443 }
2444
2445 //------------------------------construct_bb---------------------------
2446 // Construct reverse postorder list of block members
2447 bool SuperWord::construct_bb() {
2448 Node* entry = bb();
2449
2450 assert(_stk.length() == 0, "stk is empty");
2451 assert(_block.length() == 0, "block is empty");
2452 assert(_data_entry.length() == 0, "data_entry is empty");
2453 assert(_mem_slice_head.length() == 0, "mem_slice_head is empty");
2454 assert(_mem_slice_tail.length() == 0, "mem_slice_tail is empty");
2455
2456 // Find non-control nodes with no inputs from within block,
2457 // create a temporary map from node _idx to bb_idx for use
2458 // by the visited and post_visited sets,
2459 // and count number of nodes in block.
2460 int bb_ct = 0;
2461 for (uint i = 0; i < lpt()->_body.size(); i++) {
2462 Node *n = lpt()->_body.at(i);
2463 set_bb_idx(n, i); // Create a temporary map
2464 if (in_bb(n)) {
2465 if (n->is_LoadStore() || n->is_MergeMem() ||
2466 (n->is_Proj() && !n->as_Proj()->is_CFG())) {
2467 // Bailout if the loop has LoadStore, MergeMem or data Proj
2468 // nodes. Superword optimization does not work with them.
2469 return false;
2470 }
2471 bb_ct++;
2472 if (!n->is_CFG()) {
2473 bool found = false;
2474 for (uint j = 0; j < n->req(); j++) {
2475 Node* def = n->in(j);
2476 if (def && in_bb(def)) {
2477 found = true;
2478 break;
2479 }
2480 }
2481 if (!found) {
2482 assert(n != entry, "can't be entry");
2483 _data_entry.push(n);
2484 }
2485 }
2486 }
2487 }
2488
2489 // Find memory slices (head and tail)
2490 for (DUIterator_Fast imax, i = lp()->fast_outs(imax); i < imax; i++) {
2491 Node *n = lp()->fast_out(i);
2492 if (in_bb(n) && (n->is_Phi() && n->bottom_type() == Type::MEMORY)) {
2493 Node* n_tail = n->in(LoopNode::LoopBackControl);
2494 if (n_tail != n->in(LoopNode::EntryControl)) {
2495 if (!n_tail->is_Mem()) {
2496 assert(n_tail->is_Mem(), "unexpected node for memory slice: %s", n_tail->Name());
2497 return false; // Bailout
2498 }
2499 _mem_slice_head.push(n);
2500 _mem_slice_tail.push(n_tail);
2501 }
2502 }
2503 }
2504
2505 // Create an RPO list of nodes in block
2506
2507 visited_clear();
2508 post_visited_clear();
2509
2510 // Push all non-control nodes with no inputs from within block, then control entry
2511 for (int j = 0; j < _data_entry.length(); j++) {
2512 Node* n = _data_entry.at(j);
2513 visited_set(n);
2514 _stk.push(n);
2515 }
2516 visited_set(entry);
2517 _stk.push(entry);
2518
2519 // Do a depth first walk over out edges
2520 int rpo_idx = bb_ct - 1;
2521 int size;
2522 int reduction_uses = 0;
2523 while ((size = _stk.length()) > 0) {
2524 Node* n = _stk.top(); // Leave node on stack
2525 if (!visited_test_set(n)) {
2526 // forward arc in graph
2527 } else if (!post_visited_test(n)) {
2528 // cross or back arc
2529 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2530 Node *use = n->fast_out(i);
2531 if (in_bb(use) && !visited_test(use) &&
2532 // Don't go around backedge
2533 (!use->is_Phi() || n == entry)) {
2534 if (use->is_reduction()) {
2535 // First see if we can map the reduction on the given system we are on, then
2536 // make a data entry operation for each reduction we see.
2537 BasicType bt = use->bottom_type()->basic_type();
2538 if (ReductionNode::implemented(use->Opcode(), Matcher::min_vector_size(bt), bt)) {
2539 reduction_uses++;
2540 }
2541 }
2542 _stk.push(use);
2543 }
2544 }
2545 if (_stk.length() == size) {
2546 // There were no additional uses, post visit node now
2547 _stk.pop(); // Remove node from stack
2548 assert(rpo_idx >= 0, "");
2549 _block.at_put_grow(rpo_idx, n);
2550 rpo_idx--;
2551 post_visited_set(n);
2552 assert(rpo_idx >= 0 || _stk.is_empty(), "");
2553 }
2554 } else {
2555 _stk.pop(); // Remove post-visited node from stack
2556 }
2557 }//while
2558
2559 int ii_current = -1;
2560 unsigned int load_idx = (unsigned int)-1;
2561 _ii_order.clear();
2562 // Create real map of block indices for nodes
2563 for (int j = 0; j < _block.length(); j++) {
2564 Node* n = _block.at(j);
2565 set_bb_idx(n, j);
2566 if (_do_vector_loop && n->is_Load()) {
2567 if (ii_current == -1) {
2568 ii_current = _clone_map.gen(n->_idx);
2569 _ii_order.push(ii_current);
2570 load_idx = _clone_map.idx(n->_idx);
2571 } else if (_clone_map.idx(n->_idx) == load_idx && _clone_map.gen(n->_idx) != ii_current) {
2572 ii_current = _clone_map.gen(n->_idx);
2573 _ii_order.push(ii_current);
2574 }
2575 }
2576 }//for
2577
2578 // Ensure extra info is allocated.
2579 initialize_bb();
2580
2581 #ifndef PRODUCT
2582 if (_vector_loop_debug && _ii_order.length() > 0) {
2583 tty->print("SuperWord::construct_bb: List of generations: ");
2584 for (int jj = 0; jj < _ii_order.length(); ++jj) {
2585 tty->print(" %d:%d", jj, _ii_order.at(jj));
2586 }
2587 tty->print_cr(" ");
2588 }
2589 if (TraceSuperWord) {
2590 print_bb();
2591 tty->print_cr("\ndata entry nodes: %s", _data_entry.length() > 0 ? "" : "NONE");
2592 for (int m = 0; m < _data_entry.length(); m++) {
2593 tty->print("%3d ", m);
2594 _data_entry.at(m)->dump();
2595 }
2596 tty->print_cr("\nmemory slices: %s", _mem_slice_head.length() > 0 ? "" : "NONE");
2597 for (int m = 0; m < _mem_slice_head.length(); m++) {
2598 tty->print("%3d ", m); _mem_slice_head.at(m)->dump();
2599 tty->print(" "); _mem_slice_tail.at(m)->dump();
2600 }
2601 }
2602 #endif
2603 assert(rpo_idx == -1 && bb_ct == _block.length(), "all block members found");
2604 return (_mem_slice_head.length() > 0) || (reduction_uses > 0) || (_data_entry.length() > 0);
2605 }
2606
2607 //------------------------------initialize_bb---------------------------
2608 // Initialize per node info
2609 void SuperWord::initialize_bb() {
2610 Node* last = _block.at(_block.length() - 1);
2611 grow_node_info(bb_idx(last));
2612 }
2613
2614 //------------------------------bb_insert_after---------------------------
2615 // Insert n into block after pos
2616 void SuperWord::bb_insert_after(Node* n, int pos) {
2617 int n_pos = pos + 1;
2618 // Make room
2619 for (int i = _block.length() - 1; i >= n_pos; i--) {
2620 _block.at_put_grow(i+1, _block.at(i));
2621 }
2622 for (int j = _node_info.length() - 1; j >= n_pos; j--) {
2623 _node_info.at_put_grow(j+1, _node_info.at(j));
2624 }
2625 // Set value
2626 _block.at_put_grow(n_pos, n);
2627 _node_info.at_put_grow(n_pos, SWNodeInfo::initial);
2628 // Adjust map from node->_idx to _block index
2629 for (int i = n_pos; i < _block.length(); i++) {
2630 set_bb_idx(_block.at(i), i);
2631 }
2632 }
2633
2634 //------------------------------compute_max_depth---------------------------
2635 // Compute max depth for expressions from beginning of block
2636 // Use to prune search paths during test for independence.
2637 void SuperWord::compute_max_depth() {
2638 int ct = 0;
2639 bool again;
2640 do {
2641 again = false;
2642 for (int i = 0; i < _block.length(); i++) {
2643 Node* n = _block.at(i);
2644 if (!n->is_Phi()) {
2645 int d_orig = depth(n);
2646 int d_in = 0;
2647 for (DepPreds preds(n, _dg); !preds.done(); preds.next()) {
2648 Node* pred = preds.current();
2649 if (in_bb(pred)) {
2650 d_in = MAX2(d_in, depth(pred));
2651 }
2652 }
2653 if (d_in + 1 != d_orig) {
2654 set_depth(n, d_in + 1);
2655 again = true;
2656 }
2657 }
2658 }
2659 ct++;
2660 } while (again);
2661
2662 if (TraceSuperWord && Verbose) {
2663 tty->print_cr("compute_max_depth iterated: %d times", ct);
2664 }
2665 }
2666
2667 //-------------------------compute_vector_element_type-----------------------
2668 // Compute necessary vector element type for expressions
2669 // This propagates backwards a narrower integer type when the
2670 // upper bits of the value are not needed.
2671 // Example: char a,b,c; a = b + c;
2672 // Normally the type of the add is integer, but for packed character
2673 // operations the type of the add needs to be char.
2674 void SuperWord::compute_vector_element_type() {
2675 if (TraceSuperWord && Verbose) {
2676 tty->print_cr("\ncompute_velt_type:");
2677 }
2678
2679 // Initial type
2680 for (int i = 0; i < _block.length(); i++) {
2681 Node* n = _block.at(i);
2682 set_velt_type(n, container_type(n));
2683 }
2684
2685 // Propagate integer narrowed type backwards through operations
2686 // that don't depend on higher order bits
2687 for (int i = _block.length() - 1; i >= 0; i--) {
2688 Node* n = _block.at(i);
2689 // Only integer types need be examined
2690 const Type* vtn = velt_type(n);
2691 if (vtn->basic_type() == T_INT) {
2692 uint start, end;
2693 VectorNode::vector_operands(n, &start, &end);
2694
2695 for (uint j = start; j < end; j++) {
2696 Node* in = n->in(j);
2697 // Don't propagate through a memory
2698 if (!in->is_Mem() && in_bb(in) && velt_type(in)->basic_type() == T_INT &&
2699 data_size(n) < data_size(in)) {
2700 bool same_type = true;
2701 for (DUIterator_Fast kmax, k = in->fast_outs(kmax); k < kmax; k++) {
2702 Node *use = in->fast_out(k);
2703 if (!in_bb(use) || !same_velt_type(use, n)) {
2704 same_type = false;
2705 break;
2706 }
2707 }
2708 if (same_type) {
2709 // For right shifts of small integer types (bool, byte, char, short)
2710 // we need precise information about sign-ness. Only Load nodes have
2711 // this information because Store nodes are the same for signed and
2712 // unsigned values. And any arithmetic operation after a load may
2713 // expand a value to signed Int so such right shifts can't be used
2714 // because vector elements do not have upper bits of Int.
2715 const Type* vt = vtn;
2716 if (VectorNode::is_shift(in)) {
2717 Node* load = in->in(1);
2718 if (load->is_Load() && in_bb(load) && (velt_type(load)->basic_type() == T_INT)) {
2719 vt = velt_type(load);
2720 } else if (in->Opcode() != Op_LShiftI) {
2721 // Widen type to Int to avoid creation of right shift vector
2722 // (align + data_size(s1) check in stmts_can_pack() will fail).
2723 // Note, left shifts work regardless type.
2724 vt = TypeInt::INT;
2725 }
2726 }
2727 set_velt_type(in, vt);
2728 }
2729 }
2730 }
2731 }
2732 }
2733 #ifndef PRODUCT
2734 if (TraceSuperWord && Verbose) {
2735 for (int i = 0; i < _block.length(); i++) {
2736 Node* n = _block.at(i);
2737 velt_type(n)->dump();
2738 tty->print("\t");
2739 n->dump();
2740 }
2741 }
2742 #endif
2743 }
2744
2745 //------------------------------memory_alignment---------------------------
2746 // Alignment within a vector memory reference
2747 int SuperWord::memory_alignment(MemNode* s, int iv_adjust) {
2748 #ifndef PRODUCT
2749 if(TraceSuperWord && Verbose) {
2750 tty->print("SuperWord::memory_alignment within a vector memory reference for %d: ", s->_idx); s->dump();
2751 }
2752 #endif
2753 NOT_PRODUCT(SWPointer::Tracer::Depth ddd(0);)
2754 SWPointer p(s, this, NULL, false);
2755 if (!p.valid()) {
2756 NOT_PRODUCT(if(is_trace_alignment()) tty->print("SWPointer::memory_alignment: SWPointer p invalid, return bottom_align");)
2757 return bottom_align;
2758 }
2759 int vw = vector_width_in_bytes(s);
2760 if (vw < 2) {
2761 NOT_PRODUCT(if(is_trace_alignment()) tty->print_cr("SWPointer::memory_alignment: vector_width_in_bytes < 2, return bottom_align");)
2762 return bottom_align; // No vectors for this type
2763 }
2764 int offset = p.offset_in_bytes();
2765 offset += iv_adjust*p.memory_size();
2766 int off_rem = offset % vw;
2767 int off_mod = off_rem >= 0 ? off_rem : off_rem + vw;
2768 if (TraceSuperWord && Verbose) {
2769 tty->print_cr("SWPointer::memory_alignment: off_rem = %d, off_mod = %d", off_rem, off_mod);
2770 }
2771 return off_mod;
2772 }
2773
2774 //---------------------------container_type---------------------------
2775 // Smallest type containing range of values
2776 const Type* SuperWord::container_type(Node* n) {
2777 if (n->is_Mem()) {
2778 BasicType bt = n->as_Mem()->memory_type();
2779 if (n->is_Store() && (bt == T_CHAR)) {
2780 // Use T_SHORT type instead of T_CHAR for stored values because any
2781 // preceding arithmetic operation extends values to signed Int.
2782 bt = T_SHORT;
2783 }
2784 if (n->Opcode() == Op_LoadUB) {
2785 // Adjust type for unsigned byte loads, it is important for right shifts.
2786 // T_BOOLEAN is used because there is no basic type representing type
2787 // TypeInt::UBYTE. Use of T_BOOLEAN for vectors is fine because only
2788 // size (one byte) and sign is important.
2789 bt = T_BOOLEAN;
2790 }
2791 return Type::get_const_basic_type(bt);
2792 }
2793 const Type* t = _igvn.type(n);
2794 if (t->basic_type() == T_INT) {
2795 // A narrow type of arithmetic operations will be determined by
2796 // propagating the type of memory operations.
2797 return TypeInt::INT;
2798 }
2799 return t;
2800 }
2801
2802 bool SuperWord::same_velt_type(Node* n1, Node* n2) {
2803 const Type* vt1 = velt_type(n1);
2804 const Type* vt2 = velt_type(n2);
2805 if (vt1->basic_type() == T_INT && vt2->basic_type() == T_INT) {
2806 // Compare vectors element sizes for integer types.
2807 return data_size(n1) == data_size(n2);
2808 }
2809 return vt1 == vt2;
2810 }
2811
2812 //------------------------------in_packset---------------------------
2813 // Are s1 and s2 in a pack pair and ordered as s1,s2?
2814 bool SuperWord::in_packset(Node* s1, Node* s2) {
2815 for (int i = 0; i < _packset.length(); i++) {
2816 Node_List* p = _packset.at(i);
2817 assert(p->size() == 2, "must be");
2818 if (p->at(0) == s1 && p->at(p->size()-1) == s2) {
2819 return true;
2820 }
2821 }
2822 return false;
2823 }
2824
2825 //------------------------------in_pack---------------------------
2826 // Is s in pack p?
2827 Node_List* SuperWord::in_pack(Node* s, Node_List* p) {
2828 for (uint i = 0; i < p->size(); i++) {
2829 if (p->at(i) == s) {
2830 return p;
2831 }
2832 }
2833 return NULL;
2834 }
2835
2836 //------------------------------remove_pack_at---------------------------
2837 // Remove the pack at position pos in the packset
2838 void SuperWord::remove_pack_at(int pos) {
2839 Node_List* p = _packset.at(pos);
2840 for (uint i = 0; i < p->size(); i++) {
2841 Node* s = p->at(i);
2842 set_my_pack(s, NULL);
2843 }
2844 _packset.remove_at(pos);
2845 }
2846
2847 void SuperWord::packset_sort(int n) {
2848 // simple bubble sort so that we capitalize with O(n) when its already sorted
2849 while (n != 0) {
2850 bool swapped = false;
2851 for (int i = 1; i < n; i++) {
2852 Node_List* q_low = _packset.at(i-1);
2853 Node_List* q_i = _packset.at(i);
2854
2855 // only swap when we find something to swap
2856 if (alignment(q_low->at(0)) > alignment(q_i->at(0))) {
2857 Node_List* t = q_i;
2858 *(_packset.adr_at(i)) = q_low;
2859 *(_packset.adr_at(i-1)) = q_i;
2860 swapped = true;
2861 }
2862 }
2863 if (swapped == false) break;
2864 n--;
2865 }
2866 }
2867
2868 //------------------------------executed_first---------------------------
2869 // Return the node executed first in pack p. Uses the RPO block list
2870 // to determine order.
2871 Node* SuperWord::executed_first(Node_List* p) {
2872 Node* n = p->at(0);
2873 int n_rpo = bb_idx(n);
2874 for (uint i = 1; i < p->size(); i++) {
2875 Node* s = p->at(i);
2876 int s_rpo = bb_idx(s);
2877 if (s_rpo < n_rpo) {
2878 n = s;
2879 n_rpo = s_rpo;
2880 }
2881 }
2882 return n;
2883 }
2884
2885 //------------------------------executed_last---------------------------
2886 // Return the node executed last in pack p.
2887 Node* SuperWord::executed_last(Node_List* p) {
2888 Node* n = p->at(0);
2889 int n_rpo = bb_idx(n);
2890 for (uint i = 1; i < p->size(); i++) {
2891 Node* s = p->at(i);
2892 int s_rpo = bb_idx(s);
2893 if (s_rpo > n_rpo) {
2894 n = s;
2895 n_rpo = s_rpo;
2896 }
2897 }
2898 return n;
2899 }
2900
2901 LoadNode::ControlDependency SuperWord::control_dependency(Node_List* p) {
2902 LoadNode::ControlDependency dep = LoadNode::DependsOnlyOnTest;
2903 for (uint i = 0; i < p->size(); i++) {
2904 Node* n = p->at(i);
2905 assert(n->is_Load(), "only meaningful for loads");
2906 if (!n->depends_only_on_test()) {
2907 dep = LoadNode::Pinned;
2908 }
2909 }
2910 return dep;
2911 }
2912
2913
2914 //----------------------------align_initial_loop_index---------------------------
2915 // Adjust pre-loop limit so that in main loop, a load/store reference
2916 // to align_to_ref will be a position zero in the vector.
2917 // (iv + k) mod vector_align == 0
2918 void SuperWord::align_initial_loop_index(MemNode* align_to_ref) {
2919 CountedLoopNode *main_head = lp()->as_CountedLoop();
2920 assert(main_head->is_main_loop(), "");
2921 CountedLoopEndNode* pre_end = get_pre_loop_end(main_head);
2922 assert(pre_end != NULL, "we must have a correct pre-loop");
2923 Node *pre_opaq1 = pre_end->limit();
2924 assert(pre_opaq1->Opcode() == Op_Opaque1, "");
2925 Opaque1Node *pre_opaq = (Opaque1Node*)pre_opaq1;
2926 Node *lim0 = pre_opaq->in(1);
2927
2928 // Where we put new limit calculations
2929 Node *pre_ctrl = pre_end->loopnode()->in(LoopNode::EntryControl);
2930
2931 // Ensure the original loop limit is available from the
2932 // pre-loop Opaque1 node.
2933 Node *orig_limit = pre_opaq->original_loop_limit();
2934 assert(orig_limit != NULL && _igvn.type(orig_limit) != Type::TOP, "");
2935
2936 SWPointer align_to_ref_p(align_to_ref, this, NULL, false);
2937 assert(align_to_ref_p.valid(), "sanity");
2938
2939 // Given:
2940 // lim0 == original pre loop limit
2941 // V == v_align (power of 2)
2942 // invar == extra invariant piece of the address expression
2943 // e == offset [ +/- invar ]
2944 //
2945 // When reassociating expressions involving '%' the basic rules are:
2946 // (a - b) % k == 0 => a % k == b % k
2947 // and:
2948 // (a + b) % k == 0 => a % k == (k - b) % k
2949 //
2950 // For stride > 0 && scale > 0,
2951 // Derive the new pre-loop limit "lim" such that the two constraints:
2952 // (1) lim = lim0 + N (where N is some positive integer < V)
2953 // (2) (e + lim) % V == 0
2954 // are true.
2955 //
2956 // Substituting (1) into (2),
2957 // (e + lim0 + N) % V == 0
2958 // solve for N:
2959 // N = (V - (e + lim0)) % V
2960 // substitute back into (1), so that new limit
2961 // lim = lim0 + (V - (e + lim0)) % V
2962 //
2963 // For stride > 0 && scale < 0
2964 // Constraints:
2965 // lim = lim0 + N
2966 // (e - lim) % V == 0
2967 // Solving for lim:
2968 // (e - lim0 - N) % V == 0
2969 // N = (e - lim0) % V
2970 // lim = lim0 + (e - lim0) % V
2971 //
2972 // For stride < 0 && scale > 0
2973 // Constraints:
2974 // lim = lim0 - N
2975 // (e + lim) % V == 0
2976 // Solving for lim:
2977 // (e + lim0 - N) % V == 0
2978 // N = (e + lim0) % V
2979 // lim = lim0 - (e + lim0) % V
2980 //
2981 // For stride < 0 && scale < 0
2982 // Constraints:
2983 // lim = lim0 - N
2984 // (e - lim) % V == 0
2985 // Solving for lim:
2986 // (e - lim0 + N) % V == 0
2987 // N = (V - (e - lim0)) % V
2988 // lim = lim0 - (V - (e - lim0)) % V
2989
2990 int vw = vector_width_in_bytes(align_to_ref);
2991 int stride = iv_stride();
2992 int scale = align_to_ref_p.scale_in_bytes();
2993 int elt_size = align_to_ref_p.memory_size();
2994 int v_align = vw / elt_size;
2995 assert(v_align > 1, "sanity");
2996 int offset = align_to_ref_p.offset_in_bytes() / elt_size;
2997 Node *offsn = _igvn.intcon(offset);
2998
2999 Node *e = offsn;
3000 if (align_to_ref_p.invar() != NULL) {
3001 // incorporate any extra invariant piece producing (offset +/- invar) >>> log2(elt)
3002 Node* log2_elt = _igvn.intcon(exact_log2(elt_size));
3003 Node* aref = new URShiftINode(align_to_ref_p.invar(), log2_elt);
3004 _igvn.register_new_node_with_optimizer(aref);
3005 _phase->set_ctrl(aref, pre_ctrl);
3006 if (align_to_ref_p.negate_invar()) {
3007 e = new SubINode(e, aref);
3008 } else {
3009 e = new AddINode(e, aref);
3010 }
3011 _igvn.register_new_node_with_optimizer(e);
3012 _phase->set_ctrl(e, pre_ctrl);
3013 }
3014 if (vw > ObjectAlignmentInBytes) {
3015 // incorporate base e +/- base && Mask >>> log2(elt)
3016 Node* xbase = new CastP2XNode(NULL, align_to_ref_p.base());
3017 _igvn.register_new_node_with_optimizer(xbase);
3018 #ifdef _LP64
3019 xbase = new ConvL2INode(xbase);
3020 _igvn.register_new_node_with_optimizer(xbase);
3021 #endif
3022 Node* mask = _igvn.intcon(vw-1);
3023 Node* masked_xbase = new AndINode(xbase, mask);
3024 _igvn.register_new_node_with_optimizer(masked_xbase);
3025 Node* log2_elt = _igvn.intcon(exact_log2(elt_size));
3026 Node* bref = new URShiftINode(masked_xbase, log2_elt);
3027 _igvn.register_new_node_with_optimizer(bref);
3028 _phase->set_ctrl(bref, pre_ctrl);
3029 e = new AddINode(e, bref);
3030 _igvn.register_new_node_with_optimizer(e);
3031 _phase->set_ctrl(e, pre_ctrl);
3032 }
3033
3034 // compute e +/- lim0
3035 if (scale < 0) {
3036 e = new SubINode(e, lim0);
3037 } else {
3038 e = new AddINode(e, lim0);
3039 }
3040 _igvn.register_new_node_with_optimizer(e);
3041 _phase->set_ctrl(e, pre_ctrl);
3042
3043 if (stride * scale > 0) {
3044 // compute V - (e +/- lim0)
3045 Node* va = _igvn.intcon(v_align);
3046 e = new SubINode(va, e);
3047 _igvn.register_new_node_with_optimizer(e);
3048 _phase->set_ctrl(e, pre_ctrl);
3049 }
3050 // compute N = (exp) % V
3051 Node* va_msk = _igvn.intcon(v_align - 1);
3052 Node* N = new AndINode(e, va_msk);
3053 _igvn.register_new_node_with_optimizer(N);
3054 _phase->set_ctrl(N, pre_ctrl);
3055
3056 // substitute back into (1), so that new limit
3057 // lim = lim0 + N
3058 Node* lim;
3059 if (stride < 0) {
3060 lim = new SubINode(lim0, N);
3061 } else {
3062 lim = new AddINode(lim0, N);
3063 }
3064 _igvn.register_new_node_with_optimizer(lim);
3065 _phase->set_ctrl(lim, pre_ctrl);
3066 Node* constrained =
3067 (stride > 0) ? (Node*) new MinINode(lim, orig_limit)
3068 : (Node*) new MaxINode(lim, orig_limit);
3069 _igvn.register_new_node_with_optimizer(constrained);
3070 _phase->set_ctrl(constrained, pre_ctrl);
3071 _igvn.hash_delete(pre_opaq);
3072 pre_opaq->set_req(1, constrained);
3073 }
3074
3075 //----------------------------get_pre_loop_end---------------------------
3076 // Find pre loop end from main loop. Returns null if none.
3077 CountedLoopEndNode* SuperWord::get_pre_loop_end(CountedLoopNode* cl) {
3078 Node* ctrl = cl->in(LoopNode::EntryControl);
3079 if (!ctrl->is_IfTrue() && !ctrl->is_IfFalse()) return NULL;
3080 Node* iffm = ctrl->in(0);
3081 if (!iffm->is_If()) return NULL;
3082 Node* bolzm = iffm->in(1);
3083 if (!bolzm->is_Bool()) return NULL;
3084 Node* cmpzm = bolzm->in(1);
3085 if (!cmpzm->is_Cmp()) return NULL;
3086 Node* opqzm = cmpzm->in(2);
3087 // Can not optimize a loop if zero-trip Opaque1 node is optimized
3088 // away and then another round of loop opts attempted.
3089 if (opqzm->Opcode() != Op_Opaque1) {
3090 return NULL;
3091 }
3092 Node* p_f = iffm->in(0);
3093 if (!p_f->is_IfFalse()) return NULL;
3094 if (!p_f->in(0)->is_CountedLoopEnd()) return NULL;
3095 CountedLoopEndNode* pre_end = p_f->in(0)->as_CountedLoopEnd();
3096 CountedLoopNode* loop_node = pre_end->loopnode();
3097 if (loop_node == NULL || !loop_node->is_pre_loop()) return NULL;
3098 return pre_end;
3099 }
3100
3101
3102 //------------------------------init---------------------------
3103 void SuperWord::init() {
3104 _dg.init();
3105 _packset.clear();
3106 _disjoint_ptrs.clear();
3107 _block.clear();
3108 _data_entry.clear();
3109 _mem_slice_head.clear();
3110 _mem_slice_tail.clear();
3111 _iteration_first.clear();
3112 _iteration_last.clear();
3113 _node_info.clear();
3114 _align_to_ref = NULL;
3115 _lpt = NULL;
3116 _lp = NULL;
3117 _bb = NULL;
3118 _iv = NULL;
3119 _race_possible = 0;
3120 _early_return = false;
3121 _num_work_vecs = 0;
3122 _num_reductions = 0;
3123 }
3124
3125 //------------------------------restart---------------------------
3126 void SuperWord::restart() {
3127 _dg.init();
3128 _packset.clear();
3129 _disjoint_ptrs.clear();
3130 _block.clear();
3131 _data_entry.clear();
3132 _mem_slice_head.clear();
3133 _mem_slice_tail.clear();
3134 _node_info.clear();
3135 }
3136
3137 //------------------------------print_packset---------------------------
3138 void SuperWord::print_packset() {
3139 #ifndef PRODUCT
3140 tty->print_cr("packset");
3141 for (int i = 0; i < _packset.length(); i++) {
3142 tty->print_cr("Pack: %d", i);
3143 Node_List* p = _packset.at(i);
3144 print_pack(p);
3145 }
3146 #endif
3147 }
3148
3149 //------------------------------print_pack---------------------------
3150 void SuperWord::print_pack(Node_List* p) {
3151 for (uint i = 0; i < p->size(); i++) {
3152 print_stmt(p->at(i));
3153 }
3154 }
3155
3156 //------------------------------print_bb---------------------------
3157 void SuperWord::print_bb() {
3158 #ifndef PRODUCT
3159 tty->print_cr("\nBlock");
3160 for (int i = 0; i < _block.length(); i++) {
3161 Node* n = _block.at(i);
3162 tty->print("%d ", i);
3163 if (n) {
3164 n->dump();
3165 }
3166 }
3167 #endif
3168 }
3169
3170 //------------------------------print_stmt---------------------------
3171 void SuperWord::print_stmt(Node* s) {
3172 #ifndef PRODUCT
3173 tty->print(" align: %d \t", alignment(s));
3174 s->dump();
3175 #endif
3176 }
3177
3178 //------------------------------blank---------------------------
3179 char* SuperWord::blank(uint depth) {
3180 static char blanks[101];
3181 assert(depth < 101, "too deep");
3182 for (uint i = 0; i < depth; i++) blanks[i] = ' ';
3183 blanks[depth] = '\0';
3184 return blanks;
3185 }
3186
3187
3188 //==============================SWPointer===========================
3189 #ifndef PRODUCT
3190 int SWPointer::Tracer::_depth = 0;
3191 #endif
3192 //----------------------------SWPointer------------------------
3193 SWPointer::SWPointer(MemNode* mem, SuperWord* slp, Node_Stack *nstack, bool analyze_only) :
3194 _mem(mem), _slp(slp), _base(NULL), _adr(NULL),
3195 _scale(0), _offset(0), _invar(NULL), _negate_invar(false),
3196 _nstack(nstack), _analyze_only(analyze_only),
3197 _stack_idx(0)
3198 #ifndef PRODUCT
3199 , _tracer(slp)
3200 #endif
3201 {
3202 NOT_PRODUCT(_tracer.ctor_1(mem);)
3203
3204 Node* adr = mem->in(MemNode::Address);
3205 if (!adr->is_AddP()) {
3206 assert(!valid(), "too complex");
3207 return;
3208 }
3209 // Match AddP(base, AddP(ptr, k*iv [+ invariant]), constant)
3210 Node* base = adr->in(AddPNode::Base);
3211 // The base address should be loop invariant
3212 if (!invariant(base)) {
3213 assert(!valid(), "base address is loop variant");
3214 return;
3215 }
3216 //unsafe reference could not be aligned appropriately without runtime checking
3217 if (base == NULL || base->bottom_type() == Type::TOP) {
3218 assert(!valid(), "unsafe access");
3219 return;
3220 }
3221
3222 NOT_PRODUCT(if(_slp->is_trace_alignment()) _tracer.store_depth();)
3223 NOT_PRODUCT(_tracer.ctor_2(adr);)
3224
3225 int i;
3226 for (i = 0; i < 3; i++) {
3227 NOT_PRODUCT(_tracer.ctor_3(adr, i);)
3228
3229 if (!scaled_iv_plus_offset(adr->in(AddPNode::Offset))) {
3230 assert(!valid(), "too complex");
3231 return;
3232 }
3233 adr = adr->in(AddPNode::Address);
3234 NOT_PRODUCT(_tracer.ctor_4(adr, i);)
3235
3236 if (base == adr || !adr->is_AddP()) {
3237 NOT_PRODUCT(_tracer.ctor_5(adr, base, i);)
3238 break; // stop looking at addp's
3239 }
3240 }
3241 NOT_PRODUCT(if(_slp->is_trace_alignment()) _tracer.restore_depth();)
3242 NOT_PRODUCT(_tracer.ctor_6(mem);)
3243
3244 _base = base;
3245 _adr = adr;
3246 assert(valid(), "Usable");
3247 }
3248
3249 // Following is used to create a temporary object during
3250 // the pattern match of an address expression.
3251 SWPointer::SWPointer(SWPointer* p) :
3252 _mem(p->_mem), _slp(p->_slp), _base(NULL), _adr(NULL),
3253 _scale(0), _offset(0), _invar(NULL), _negate_invar(false),
3254 _nstack(p->_nstack), _analyze_only(p->_analyze_only),
3255 _stack_idx(p->_stack_idx)
3256 #ifndef PRODUCT
3257 , _tracer(p->_slp)
3258 #endif
3259 {}
3260
3261
3262 bool SWPointer::invariant(Node* n) {
3263 NOT_PRODUCT(Tracer::Depth dd;)
3264 Node *n_c = phase()->get_ctrl(n);
3265 NOT_PRODUCT(_tracer.invariant_1(n, n_c);)
3266 return !lpt()->is_member(phase()->get_loop(n_c));
3267 }
3268 //------------------------scaled_iv_plus_offset--------------------
3269 // Match: k*iv + offset
3270 // where: k is a constant that maybe zero, and
3271 // offset is (k2 [+/- invariant]) where k2 maybe zero and invariant is optional
3272 bool SWPointer::scaled_iv_plus_offset(Node* n) {
3273 NOT_PRODUCT(Tracer::Depth ddd;)
3274 NOT_PRODUCT(_tracer.scaled_iv_plus_offset_1(n);)
3275
3276 if (scaled_iv(n)) {
3277 NOT_PRODUCT(_tracer.scaled_iv_plus_offset_2(n);)
3278 return true;
3279 }
3280
3281 if (offset_plus_k(n)) {
3282 NOT_PRODUCT(_tracer.scaled_iv_plus_offset_3(n);)
3283 return true;
3284 }
3285
3286 int opc = n->Opcode();
3287 if (opc == Op_AddI) {
3288 if (scaled_iv(n->in(1)) && offset_plus_k(n->in(2))) {
3289 NOT_PRODUCT(_tracer.scaled_iv_plus_offset_4(n);)
3290 return true;
3291 }
3292 if (scaled_iv(n->in(2)) && offset_plus_k(n->in(1))) {
3293 NOT_PRODUCT(_tracer.scaled_iv_plus_offset_5(n);)
3294 return true;
3295 }
3296 } else if (opc == Op_SubI) {
3297 if (scaled_iv(n->in(1)) && offset_plus_k(n->in(2), true)) {
3298 NOT_PRODUCT(_tracer.scaled_iv_plus_offset_6(n);)
3299 return true;
3300 }
3301 if (scaled_iv(n->in(2)) && offset_plus_k(n->in(1))) {
3302 _scale *= -1;
3303 NOT_PRODUCT(_tracer.scaled_iv_plus_offset_7(n);)
3304 return true;
3305 }
3306 }
3307
3308 NOT_PRODUCT(_tracer.scaled_iv_plus_offset_8(n);)
3309 return false;
3310 }
3311
3312 //----------------------------scaled_iv------------------------
3313 // Match: k*iv where k is a constant that's not zero
3314 bool SWPointer::scaled_iv(Node* n) {
3315 NOT_PRODUCT(Tracer::Depth ddd;)
3316 NOT_PRODUCT(_tracer.scaled_iv_1(n);)
3317
3318 if (_scale != 0) { // already found a scale
3319 NOT_PRODUCT(_tracer.scaled_iv_2(n, _scale);)
3320 return false;
3321 }
3322
3323 if (n == iv()) {
3324 _scale = 1;
3325 NOT_PRODUCT(_tracer.scaled_iv_3(n, _scale);)
3326 return true;
3327 }
3328 if (_analyze_only && (invariant(n) == false)) {
3329 _nstack->push(n, _stack_idx++);
3330 }
3331
3332 int opc = n->Opcode();
3333 if (opc == Op_MulI) {
3334 if (n->in(1) == iv() && n->in(2)->is_Con()) {
3335 _scale = n->in(2)->get_int();
3336 NOT_PRODUCT(_tracer.scaled_iv_4(n, _scale);)
3337 return true;
3338 } else if (n->in(2) == iv() && n->in(1)->is_Con()) {
3339 _scale = n->in(1)->get_int();
3340 NOT_PRODUCT(_tracer.scaled_iv_5(n, _scale);)
3341 return true;
3342 }
3343 } else if (opc == Op_LShiftI) {
3344 if (n->in(1) == iv() && n->in(2)->is_Con()) {
3345 _scale = 1 << n->in(2)->get_int();
3346 NOT_PRODUCT(_tracer.scaled_iv_6(n, _scale);)
3347 return true;
3348 }
3349 } else if (opc == Op_ConvI2L) {
3350 if (n->in(1)->Opcode() == Op_CastII &&
3351 n->in(1)->as_CastII()->has_range_check()) {
3352 // Skip range check dependent CastII nodes
3353 n = n->in(1);
3354 }
3355 if (scaled_iv_plus_offset(n->in(1))) {
3356 NOT_PRODUCT(_tracer.scaled_iv_7(n);)
3357 return true;
3358 }
3359 } else if (opc == Op_LShiftL) {
3360 if (!has_iv() && _invar == NULL) {
3361 // Need to preserve the current _offset value, so
3362 // create a temporary object for this expression subtree.
3363 // Hacky, so should re-engineer the address pattern match.
3364 NOT_PRODUCT(Tracer::Depth dddd;)
3365 SWPointer tmp(this);
3366 NOT_PRODUCT(_tracer.scaled_iv_8(n, &tmp);)
3367
3368 if (tmp.scaled_iv_plus_offset(n->in(1))) {
3369 if (tmp._invar == NULL || _slp->do_vector_loop()) {
3370 int mult = 1 << n->in(2)->get_int();
3371 _scale = tmp._scale * mult;
3372 _offset += tmp._offset * mult;
3373 NOT_PRODUCT(_tracer.scaled_iv_9(n, _scale, _offset, mult);)
3374 return true;
3375 }
3376 }
3377 }
3378 }
3379 NOT_PRODUCT(_tracer.scaled_iv_10(n);)
3380 return false;
3381 }
3382
3383 //----------------------------offset_plus_k------------------------
3384 // Match: offset is (k [+/- invariant])
3385 // where k maybe zero and invariant is optional, but not both.
3386 bool SWPointer::offset_plus_k(Node* n, bool negate) {
3387 NOT_PRODUCT(Tracer::Depth ddd;)
3388 NOT_PRODUCT(_tracer.offset_plus_k_1(n);)
3389
3390 int opc = n->Opcode();
3391 if (opc == Op_ConI) {
3392 _offset += negate ? -(n->get_int()) : n->get_int();
3393 NOT_PRODUCT(_tracer.offset_plus_k_2(n, _offset);)
3394 return true;
3395 } else if (opc == Op_ConL) {
3396 // Okay if value fits into an int
3397 const TypeLong* t = n->find_long_type();
3398 if (t->higher_equal(TypeLong::INT)) {
3399 jlong loff = n->get_long();
3400 jint off = (jint)loff;
3401 _offset += negate ? -off : loff;
3402 NOT_PRODUCT(_tracer.offset_plus_k_3(n, _offset);)
3403 return true;
3404 }
3405 NOT_PRODUCT(_tracer.offset_plus_k_4(n);)
3406 return false;
3407 }
3408 if (_invar != NULL) { // already has an invariant
3409 NOT_PRODUCT(_tracer.offset_plus_k_5(n, _invar);)
3410 return false;
3411 }
3412
3413 if (_analyze_only && (invariant(n) == false)) {
3414 _nstack->push(n, _stack_idx++);
3415 }
3416 if (opc == Op_AddI) {
3417 if (n->in(2)->is_Con() && invariant(n->in(1))) {
3418 _negate_invar = negate;
3419 _invar = n->in(1);
3420 _offset += negate ? -(n->in(2)->get_int()) : n->in(2)->get_int();
3421 NOT_PRODUCT(_tracer.offset_plus_k_6(n, _invar, _negate_invar, _offset);)
3422 return true;
3423 } else if (n->in(1)->is_Con() && invariant(n->in(2))) {
3424 _offset += negate ? -(n->in(1)->get_int()) : n->in(1)->get_int();
3425 _negate_invar = negate;
3426 _invar = n->in(2);
3427 NOT_PRODUCT(_tracer.offset_plus_k_7(n, _invar, _negate_invar, _offset);)
3428 return true;
3429 }
3430 }
3431 if (opc == Op_SubI) {
3432 if (n->in(2)->is_Con() && invariant(n->in(1))) {
3433 _negate_invar = negate;
3434 _invar = n->in(1);
3435 _offset += !negate ? -(n->in(2)->get_int()) : n->in(2)->get_int();
3436 NOT_PRODUCT(_tracer.offset_plus_k_8(n, _invar, _negate_invar, _offset);)
3437 return true;
3438 } else if (n->in(1)->is_Con() && invariant(n->in(2))) {
3439 _offset += negate ? -(n->in(1)->get_int()) : n->in(1)->get_int();
3440 _negate_invar = !negate;
3441 _invar = n->in(2);
3442 NOT_PRODUCT(_tracer.offset_plus_k_9(n, _invar, _negate_invar, _offset);)
3443 return true;
3444 }
3445 }
3446 if (invariant(n)) {
3447 if (opc == Op_ConvI2L) {
3448 n = n->in(1);
3449 if (n->Opcode() == Op_CastII &&
3450 n->as_CastII()->has_range_check()) {
3451 // Skip range check dependent CastII nodes
3452 assert(invariant(n), "sanity");
3453 n = n->in(1);
3454 }
3455 }
3456 if (n->bottom_type()->isa_int()) {
3457 _negate_invar = negate;
3458 _invar = n;
3459 NOT_PRODUCT(_tracer.offset_plus_k_10(n, _invar, _negate_invar, _offset);)
3460 return true;
3461 }
3462 }
3463
3464 NOT_PRODUCT(_tracer.offset_plus_k_11(n);)
3465 return false;
3466 }
3467
3468 //----------------------------print------------------------
3469 void SWPointer::print() {
3470 #ifndef PRODUCT
3471 tty->print("base: %d adr: %d scale: %d offset: %d invar: %c%d\n",
3472 _base != NULL ? _base->_idx : 0,
3473 _adr != NULL ? _adr->_idx : 0,
3474 _scale, _offset,
3475 _negate_invar?'-':'+',
3476 _invar != NULL ? _invar->_idx : 0);
3477 #endif
3478 }
3479
3480 //----------------------------tracing------------------------
3481 #ifndef PRODUCT
3482 void SWPointer::Tracer::print_depth() {
3483 for (int ii = 0; ii<_depth; ++ii) tty->print(" ");
3484 }
3485
3486 void SWPointer::Tracer::ctor_1 (Node* mem) {
3487 if(_slp->is_trace_alignment()) {
3488 print_depth(); tty->print(" %d SWPointer::SWPointer: start alignment analysis", mem->_idx); mem->dump();
3489 }
3490 }
3491
3492 void SWPointer::Tracer::ctor_2(Node* adr) {
3493 if(_slp->is_trace_alignment()) {
3494 //store_depth();
3495 inc_depth();
3496 print_depth(); tty->print(" %d (adr) SWPointer::SWPointer: ", adr->_idx); adr->dump();
3497 inc_depth();
3498 print_depth(); tty->print(" %d (base) SWPointer::SWPointer: ", adr->in(AddPNode::Base)->_idx); adr->in(AddPNode::Base)->dump();
3499 }
3500 }
3501
3502 void SWPointer::Tracer::ctor_3(Node* adr, int i) {
3503 if(_slp->is_trace_alignment()) {
3504 inc_depth();
3505 Node* offset = adr->in(AddPNode::Offset);
3506 print_depth(); tty->print(" %d (offset) SWPointer::SWPointer: i = %d: ", offset->_idx, i); offset->dump();
3507 }
3508 }
3509
3510 void SWPointer::Tracer::ctor_4(Node* adr, int i) {
3511 if(_slp->is_trace_alignment()) {
3512 inc_depth();
3513 print_depth(); tty->print(" %d (adr) SWPointer::SWPointer: i = %d: ", adr->_idx, i); adr->dump();
3514 }
3515 }
3516
3517 void SWPointer::Tracer::ctor_5(Node* adr, Node* base, int i) {
3518 if(_slp->is_trace_alignment()) {
3519 inc_depth();
3520 if (base == adr) {
3521 print_depth(); tty->print_cr(" \\ %d (adr) == %d (base) SWPointer::SWPointer: breaking analysis at i = %d", adr->_idx, base->_idx, i);
3522 } else if (!adr->is_AddP()) {
3523 print_depth(); tty->print_cr(" \\ %d (adr) is NOT Addp SWPointer::SWPointer: breaking analysis at i = %d", adr->_idx, i);
3524 }
3525 }
3526 }
3527
3528 void SWPointer::Tracer::ctor_6(Node* mem) {
3529 if(_slp->is_trace_alignment()) {
3530 //restore_depth();
3531 print_depth(); tty->print_cr(" %d (adr) SWPointer::SWPointer: stop analysis", mem->_idx);
3532 }
3533 }
3534
3535 void SWPointer::Tracer::invariant_1(Node *n, Node *n_c) {
3536 if (_slp->do_vector_loop() && _slp->is_debug() && _slp->_lpt->is_member(_slp->_phase->get_loop(n_c)) != (int)_slp->in_bb(n)) {
3537 int is_member = _slp->_lpt->is_member(_slp->_phase->get_loop(n_c));
3538 int in_bb = _slp->in_bb(n);
3539 print_depth(); tty->print(" \\ "); tty->print_cr(" %d SWPointer::invariant conditions differ: n_c %d", n->_idx, n_c->_idx);
3540 print_depth(); tty->print(" \\ "); tty->print_cr("is_member %d, in_bb %d", is_member, in_bb);
3541 print_depth(); tty->print(" \\ "); n->dump();
3542 print_depth(); tty->print(" \\ "); n_c->dump();
3543 }
3544 }
3545
3546 void SWPointer::Tracer::scaled_iv_plus_offset_1(Node* n) {
3547 if(_slp->is_trace_alignment()) {
3548 print_depth(); tty->print(" %d SWPointer::scaled_iv_plus_offset testing node: ", n->_idx);
3549 n->dump();
3550 }
3551 }
3552
3553 void SWPointer::Tracer::scaled_iv_plus_offset_2(Node* n) {
3554 if(_slp->is_trace_alignment()) {
3555 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: PASSED", n->_idx);
3556 }
3557 }
3558
3559 void SWPointer::Tracer::scaled_iv_plus_offset_3(Node* n) {
3560 if(_slp->is_trace_alignment()) {
3561 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: PASSED", n->_idx);
3562 }
3563 }
3564
3565 void SWPointer::Tracer::scaled_iv_plus_offset_4(Node* n) {
3566 if(_slp->is_trace_alignment()) {
3567 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: Op_AddI PASSED", n->_idx);
3568 print_depth(); tty->print(" \\ %d SWPointer::scaled_iv_plus_offset: in(1) is scaled_iv: ", n->in(1)->_idx); n->in(1)->dump();
3569 print_depth(); tty->print(" \\ %d SWPointer::scaled_iv_plus_offset: in(2) is offset_plus_k: ", n->in(2)->_idx); n->in(2)->dump();
3570 }
3571 }
3572
3573 void SWPointer::Tracer::scaled_iv_plus_offset_5(Node* n) {
3574 if(_slp->is_trace_alignment()) {
3575 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: Op_AddI PASSED", n->_idx);
3576 print_depth(); tty->print(" \\ %d SWPointer::scaled_iv_plus_offset: in(2) is scaled_iv: ", n->in(2)->_idx); n->in(2)->dump();
3577 print_depth(); tty->print(" \\ %d SWPointer::scaled_iv_plus_offset: in(1) is offset_plus_k: ", n->in(1)->_idx); n->in(1)->dump();
3578 }
3579 }
3580
3581 void SWPointer::Tracer::scaled_iv_plus_offset_6(Node* n) {
3582 if(_slp->is_trace_alignment()) {
3583 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: Op_SubI PASSED", n->_idx);
3584 print_depth(); tty->print(" \\ %d SWPointer::scaled_iv_plus_offset: in(1) is scaled_iv: ", n->in(1)->_idx); n->in(1)->dump();
3585 print_depth(); tty->print(" \\ %d SWPointer::scaled_iv_plus_offset: in(2) is offset_plus_k: ", n->in(2)->_idx); n->in(2)->dump();
3586 }
3587 }
3588
3589 void SWPointer::Tracer::scaled_iv_plus_offset_7(Node* n) {
3590 if(_slp->is_trace_alignment()) {
3591 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: Op_SubI PASSED", n->_idx);
3592 print_depth(); tty->print(" \\ %d SWPointer::scaled_iv_plus_offset: in(2) is scaled_iv: ", n->in(2)->_idx); n->in(2)->dump();
3593 print_depth(); tty->print(" \\ %d SWPointer::scaled_iv_plus_offset: in(1) is offset_plus_k: ", n->in(1)->_idx); n->in(1)->dump();
3594 }
3595 }
3596
3597 void SWPointer::Tracer::scaled_iv_plus_offset_8(Node* n) {
3598 if(_slp->is_trace_alignment()) {
3599 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: FAILED", n->_idx);
3600 }
3601 }
3602
3603 void SWPointer::Tracer::scaled_iv_1(Node* n) {
3604 if(_slp->is_trace_alignment()) {
3605 print_depth(); tty->print(" %d SWPointer::scaled_iv: testing node: ", n->_idx); n->dump();
3606 }
3607 }
3608
3609 void SWPointer::Tracer::scaled_iv_2(Node* n, int scale) {
3610 if(_slp->is_trace_alignment()) {
3611 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: FAILED since another _scale has been detected before", n->_idx);
3612 print_depth(); tty->print_cr(" \\ SWPointer::scaled_iv: _scale (%d) != 0", scale);
3613 }
3614 }
3615
3616 void SWPointer::Tracer::scaled_iv_3(Node* n, int scale) {
3617 if(_slp->is_trace_alignment()) {
3618 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: is iv, setting _scale = %d", n->_idx, scale);
3619 }
3620 }
3621
3622 void SWPointer::Tracer::scaled_iv_4(Node* n, int scale) {
3623 if(_slp->is_trace_alignment()) {
3624 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: Op_MulI PASSED, setting _scale = %d", n->_idx, scale);
3625 print_depth(); tty->print(" \\ %d SWPointer::scaled_iv: in(1) is iv: ", n->in(1)->_idx); n->in(1)->dump();
3626 print_depth(); tty->print(" \\ %d SWPointer::scaled_iv: in(2) is Con: ", n->in(2)->_idx); n->in(2)->dump();
3627 }
3628 }
3629
3630 void SWPointer::Tracer::scaled_iv_5(Node* n, int scale) {
3631 if(_slp->is_trace_alignment()) {
3632 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: Op_MulI PASSED, setting _scale = %d", n->_idx, scale);
3633 print_depth(); tty->print(" \\ %d SWPointer::scaled_iv: in(2) is iv: ", n->in(2)->_idx); n->in(2)->dump();
3634 print_depth(); tty->print(" \\ %d SWPointer::scaled_iv: in(1) is Con: ", n->in(1)->_idx); n->in(1)->dump();
3635 }
3636 }
3637
3638 void SWPointer::Tracer::scaled_iv_6(Node* n, int scale) {
3639 if(_slp->is_trace_alignment()) {
3640 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: Op_LShiftI PASSED, setting _scale = %d", n->_idx, scale);
3641 print_depth(); tty->print(" \\ %d SWPointer::scaled_iv: in(1) is iv: ", n->in(1)->_idx); n->in(1)->dump();
3642 print_depth(); tty->print(" \\ %d SWPointer::scaled_iv: in(2) is Con: ", n->in(2)->_idx); n->in(2)->dump();
3643 }
3644 }
3645
3646 void SWPointer::Tracer::scaled_iv_7(Node* n) {
3647 if(_slp->is_trace_alignment()) {
3648 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: Op_ConvI2L PASSED", n->_idx);
3649 print_depth(); tty->print_cr(" \\ SWPointer::scaled_iv: in(1) %d is scaled_iv_plus_offset: ", n->in(1)->_idx);
3650 inc_depth(); inc_depth();
3651 print_depth(); n->in(1)->dump();
3652 dec_depth(); dec_depth();
3653 }
3654 }
3655
3656 void SWPointer::Tracer::scaled_iv_8(Node* n, SWPointer* tmp) {
3657 if(_slp->is_trace_alignment()) {
3658 print_depth(); tty->print(" %d SWPointer::scaled_iv: Op_LShiftL, creating tmp SWPointer: ", n->_idx); tmp->print();
3659 }
3660 }
3661
3662 void SWPointer::Tracer::scaled_iv_9(Node* n, int scale, int _offset, int mult) {
3663 if(_slp->is_trace_alignment()) {
3664 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: Op_LShiftL PASSED, setting _scale = %d, _offset = %d", n->_idx, scale, _offset);
3665 print_depth(); tty->print_cr(" \\ SWPointer::scaled_iv: in(1) %d is scaled_iv_plus_offset, in(2) %d used to get mult = %d: _scale = %d, _offset = %d",
3666 n->in(1)->_idx, n->in(2)->_idx, mult, scale, _offset);
3667 inc_depth(); inc_depth();
3668 print_depth(); n->in(1)->dump();
3669 print_depth(); n->in(2)->dump();
3670 dec_depth(); dec_depth();
3671 }
3672 }
3673
3674 void SWPointer::Tracer::scaled_iv_10(Node* n) {
3675 if(_slp->is_trace_alignment()) {
3676 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: FAILED", n->_idx);
3677 }
3678 }
3679
3680 void SWPointer::Tracer::offset_plus_k_1(Node* n) {
3681 if(_slp->is_trace_alignment()) {
3682 print_depth(); tty->print(" %d SWPointer::offset_plus_k: testing node: ", n->_idx); n->dump();
3683 }
3684 }
3685
3686 void SWPointer::Tracer::offset_plus_k_2(Node* n, int _offset) {
3687 if(_slp->is_trace_alignment()) {
3688 print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: Op_ConI PASSED, setting _offset = %d", n->_idx, _offset);
3689 }
3690 }
3691
3692 void SWPointer::Tracer::offset_plus_k_3(Node* n, int _offset) {
3693 if(_slp->is_trace_alignment()) {
3694 print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: Op_ConL PASSED, setting _offset = %d", n->_idx, _offset);
3695 }
3696 }
3697
3698 void SWPointer::Tracer::offset_plus_k_4(Node* n) {
3699 if(_slp->is_trace_alignment()) {
3700 print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: FAILED", n->_idx);
3701 print_depth(); tty->print_cr(" \\ " JLONG_FORMAT " SWPointer::offset_plus_k: Op_ConL FAILED, k is too big", n->get_long());
3702 }
3703 }
3704
3705 void SWPointer::Tracer::offset_plus_k_5(Node* n, Node* _invar) {
3706 if(_slp->is_trace_alignment()) {
3707 print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: FAILED since another invariant has been detected before", n->_idx);
3708 print_depth(); tty->print(" \\ %d SWPointer::offset_plus_k: _invar != NULL: ", _invar->_idx); _invar->dump();
3709 }
3710 }
3711
3712 void SWPointer::Tracer::offset_plus_k_6(Node* n, Node* _invar, bool _negate_invar, int _offset) {
3713 if(_slp->is_trace_alignment()) {
3714 print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: Op_AddI PASSED, setting _negate_invar = %d, _invar = %d, _offset = %d",
3715 n->_idx, _negate_invar, _invar->_idx, _offset);
3716 print_depth(); tty->print(" \\ %d SWPointer::offset_plus_k: in(2) is Con: ", n->in(2)->_idx); n->in(2)->dump();
3717 print_depth(); tty->print(" \\ %d SWPointer::offset_plus_k: in(1) is invariant: ", _invar->_idx); _invar->dump();
3718 }
3719 }
3720
3721 void SWPointer::Tracer::offset_plus_k_7(Node* n, Node* _invar, bool _negate_invar, int _offset) {
3722 if(_slp->is_trace_alignment()) {
3723 print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: Op_AddI PASSED, setting _negate_invar = %d, _invar = %d, _offset = %d",
3724 n->_idx, _negate_invar, _invar->_idx, _offset);
3725 print_depth(); tty->print(" \\ %d SWPointer::offset_plus_k: in(1) is Con: ", n->in(1)->_idx); n->in(1)->dump();
3726 print_depth(); tty->print(" \\ %d SWPointer::offset_plus_k: in(2) is invariant: ", _invar->_idx); _invar->dump();
3727 }
3728 }
3729
3730 void SWPointer::Tracer::offset_plus_k_8(Node* n, Node* _invar, bool _negate_invar, int _offset) {
3731 if(_slp->is_trace_alignment()) {
3732 print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: Op_SubI is PASSED, setting _negate_invar = %d, _invar = %d, _offset = %d",
3733 n->_idx, _negate_invar, _invar->_idx, _offset);
3734 print_depth(); tty->print(" \\ %d SWPointer::offset_plus_k: in(2) is Con: ", n->in(2)->_idx); n->in(2)->dump();
3735 print_depth(); tty->print(" \\ %d SWPointer::offset_plus_k: in(1) is invariant: ", _invar->_idx); _invar->dump();
3736 }
3737 }
3738
3739 void SWPointer::Tracer::offset_plus_k_9(Node* n, Node* _invar, bool _negate_invar, int _offset) {
3740 if(_slp->is_trace_alignment()) {
3741 print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: Op_SubI PASSED, setting _negate_invar = %d, _invar = %d, _offset = %d", n->_idx, _negate_invar, _invar->_idx, _offset);
3742 print_depth(); tty->print(" \\ %d SWPointer::offset_plus_k: in(1) is Con: ", n->in(1)->_idx); n->in(1)->dump();
3743 print_depth(); tty->print(" \\ %d SWPointer::offset_plus_k: in(2) is invariant: ", _invar->_idx); _invar->dump();
3744 }
3745 }
3746
3747 void SWPointer::Tracer::offset_plus_k_10(Node* n, Node* _invar, bool _negate_invar, int _offset) {
3748 if(_slp->is_trace_alignment()) {
3749 print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: PASSED, setting _negate_invar = %d, _invar = %d, _offset = %d", n->_idx, _negate_invar, _invar->_idx, _offset);
3750 print_depth(); tty->print_cr(" \\ %d SWPointer::offset_plus_k: is invariant", n->_idx);
3751 }
3752 }
3753
3754 void SWPointer::Tracer::offset_plus_k_11(Node* n) {
3755 if(_slp->is_trace_alignment()) {
3756 print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: FAILED", n->_idx);
3757 }
3758 }
3759
3760 #endif
3761 // ========================= OrderedPair =====================
3762
3763 const OrderedPair OrderedPair::initial;
3764
3765 // ========================= SWNodeInfo =====================
3766
3767 const SWNodeInfo SWNodeInfo::initial;
3768
3769
3770 // ============================ DepGraph ===========================
3771
3772 //------------------------------make_node---------------------------
3773 // Make a new dependence graph node for an ideal node.
3774 DepMem* DepGraph::make_node(Node* node) {
3775 DepMem* m = new (_arena) DepMem(node);
3776 if (node != NULL) {
3777 assert(_map.at_grow(node->_idx) == NULL, "one init only");
3778 _map.at_put_grow(node->_idx, m);
3779 }
3780 return m;
3781 }
3782
3783 //------------------------------make_edge---------------------------
3784 // Make a new dependence graph edge from dpred -> dsucc
3785 DepEdge* DepGraph::make_edge(DepMem* dpred, DepMem* dsucc) {
3786 DepEdge* e = new (_arena) DepEdge(dpred, dsucc, dsucc->in_head(), dpred->out_head());
3787 dpred->set_out_head(e);
3788 dsucc->set_in_head(e);
3789 return e;
3790 }
3791
3792 // ========================== DepMem ========================
3793
3794 //------------------------------in_cnt---------------------------
3795 int DepMem::in_cnt() {
3796 int ct = 0;
3797 for (DepEdge* e = _in_head; e != NULL; e = e->next_in()) ct++;
3798 return ct;
3799 }
3800
3801 //------------------------------out_cnt---------------------------
3802 int DepMem::out_cnt() {
3803 int ct = 0;
3804 for (DepEdge* e = _out_head; e != NULL; e = e->next_out()) ct++;
3805 return ct;
3806 }
3807
3808 //------------------------------print-----------------------------
3809 void DepMem::print() {
3810 #ifndef PRODUCT
3811 tty->print(" DepNode %d (", _node->_idx);
3812 for (DepEdge* p = _in_head; p != NULL; p = p->next_in()) {
3813 Node* pred = p->pred()->node();
3814 tty->print(" %d", pred != NULL ? pred->_idx : 0);
3815 }
3816 tty->print(") [");
3817 for (DepEdge* s = _out_head; s != NULL; s = s->next_out()) {
3818 Node* succ = s->succ()->node();
3819 tty->print(" %d", succ != NULL ? succ->_idx : 0);
3820 }
3821 tty->print_cr(" ]");
3822 #endif
3823 }
3824
3825 // =========================== DepEdge =========================
3826
3827 //------------------------------DepPreds---------------------------
3828 void DepEdge::print() {
3829 #ifndef PRODUCT
3830 tty->print_cr("DepEdge: %d [ %d ]", _pred->node()->_idx, _succ->node()->_idx);
3831 #endif
3832 }
3833
3834 // =========================== DepPreds =========================
3835 // Iterator over predecessor edges in the dependence graph.
3836
3837 //------------------------------DepPreds---------------------------
3838 DepPreds::DepPreds(Node* n, DepGraph& dg) {
3839 _n = n;
3840 _done = false;
3841 if (_n->is_Store() || _n->is_Load()) {
3842 _next_idx = MemNode::Address;
3843 _end_idx = n->req();
3844 _dep_next = dg.dep(_n)->in_head();
3845 } else if (_n->is_Mem()) {
3846 _next_idx = 0;
3847 _end_idx = 0;
3848 _dep_next = dg.dep(_n)->in_head();
3849 } else {
3850 _next_idx = 1;
3851 _end_idx = _n->req();
3852 _dep_next = NULL;
3853 }
3854 next();
3855 }
3856
3857 //------------------------------next---------------------------
3858 void DepPreds::next() {
3859 if (_dep_next != NULL) {
3860 _current = _dep_next->pred()->node();
3861 _dep_next = _dep_next->next_in();
3862 } else if (_next_idx < _end_idx) {
3863 _current = _n->in(_next_idx++);
3864 } else {
3865 _done = true;
3866 }
3867 }
3868
3869 // =========================== DepSuccs =========================
3870 // Iterator over successor edges in the dependence graph.
3871
3872 //------------------------------DepSuccs---------------------------
3873 DepSuccs::DepSuccs(Node* n, DepGraph& dg) {
3874 _n = n;
3875 _done = false;
3876 if (_n->is_Load()) {
3877 _next_idx = 0;
3878 _end_idx = _n->outcnt();
3879 _dep_next = dg.dep(_n)->out_head();
3880 } else if (_n->is_Mem() || _n->is_Phi() && _n->bottom_type() == Type::MEMORY) {
3881 _next_idx = 0;
3882 _end_idx = 0;
3883 _dep_next = dg.dep(_n)->out_head();
3884 } else {
3885 _next_idx = 0;
3886 _end_idx = _n->outcnt();
3887 _dep_next = NULL;
3888 }
3889 next();
3890 }
3891
3892 //-------------------------------next---------------------------
3893 void DepSuccs::next() {
3894 if (_dep_next != NULL) {
3895 _current = _dep_next->succ()->node();
3896 _dep_next = _dep_next->next_out();
3897 } else if (_next_idx < _end_idx) {
3898 _current = _n->raw_out(_next_idx++);
3899 } else {
3900 _done = true;
3901 }
3902 }
3903
3904 //
3905 // --------------------------------- vectorization/simd -----------------------------------
3906 //
3907 bool SuperWord::same_origin_idx(Node* a, Node* b) const {
3908 return a != NULL && b != NULL && _clone_map.same_idx(a->_idx, b->_idx);
3909 }
3910 bool SuperWord::same_generation(Node* a, Node* b) const {
3911 return a != NULL && b != NULL && _clone_map.same_gen(a->_idx, b->_idx);
3912 }
3913
3914 Node* SuperWord::find_phi_for_mem_dep(LoadNode* ld) {
3915 assert(in_bb(ld), "must be in block");
3916 if (_clone_map.gen(ld->_idx) == _ii_first) {
3917 #ifndef PRODUCT
3918 if (_vector_loop_debug) {
3919 tty->print_cr("SuperWord::find_phi_for_mem_dep _clone_map.gen(ld->_idx)=%d",
3920 _clone_map.gen(ld->_idx));
3921 }
3922 #endif
3923 return NULL; //we think that any ld in the first gen being vectorizable
3924 }
3925
3926 Node* mem = ld->in(MemNode::Memory);
3927 if (mem->outcnt() <= 1) {
3928 // we don't want to remove the only edge from mem node to load
3929 #ifndef PRODUCT
3930 if (_vector_loop_debug) {
3931 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",
3932 mem->_idx, ld->_idx);
3933 ld->dump();
3934 mem->dump();
3935 }
3936 #endif
3937 return NULL;
3938 }
3939 if (!in_bb(mem) || same_generation(mem, ld)) {
3940 #ifndef PRODUCT
3941 if (_vector_loop_debug) {
3942 tty->print_cr("SuperWord::find_phi_for_mem_dep _clone_map.gen(mem->_idx)=%d",
3943 _clone_map.gen(mem->_idx));
3944 }
3945 #endif
3946 return NULL; // does not depend on loop volatile node or depends on the same generation
3947 }
3948
3949 //otherwise first node should depend on mem-phi
3950 Node* first = first_node(ld);
3951 assert(first->is_Load(), "must be Load");
3952 Node* phi = first->as_Load()->in(MemNode::Memory);
3953 if (!phi->is_Phi() || phi->bottom_type() != Type::MEMORY) {
3954 #ifndef PRODUCT
3955 if (_vector_loop_debug) {
3956 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");
3957 ld->dump();
3958 first->dump();
3959 }
3960 #endif
3961 return NULL;
3962 }
3963
3964 Node* tail = 0;
3965 for (int m = 0; m < _mem_slice_head.length(); m++) {
3966 if (_mem_slice_head.at(m) == phi) {
3967 tail = _mem_slice_tail.at(m);
3968 }
3969 }
3970 if (tail == 0) { //test that found phi is in the list _mem_slice_head
3971 #ifndef PRODUCT
3972 if (_vector_loop_debug) {
3973 tty->print_cr("SuperWord::find_phi_for_mem_dep load %d is not vectorizable node, its phi %d is not _mem_slice_head",
3974 ld->_idx, phi->_idx);
3975 ld->dump();
3976 phi->dump();
3977 }
3978 #endif
3979 return NULL;
3980 }
3981
3982 // now all conditions are met
3983 return phi;
3984 }
3985
3986 Node* SuperWord::first_node(Node* nd) {
3987 for (int ii = 0; ii < _iteration_first.length(); ii++) {
3988 Node* nnn = _iteration_first.at(ii);
3989 if (same_origin_idx(nnn, nd)) {
3990 #ifndef PRODUCT
3991 if (_vector_loop_debug) {
3992 tty->print_cr("SuperWord::first_node: %d is the first iteration node for %d (_clone_map.idx(nnn->_idx) = %d)",
3993 nnn->_idx, nd->_idx, _clone_map.idx(nnn->_idx));
3994 }
3995 #endif
3996 return nnn;
3997 }
3998 }
3999
4000 #ifndef PRODUCT
4001 if (_vector_loop_debug) {
4002 tty->print_cr("SuperWord::first_node: did not find first iteration node for %d (_clone_map.idx(nd->_idx)=%d)",
4003 nd->_idx, _clone_map.idx(nd->_idx));
4004 }
4005 #endif
4006 return 0;
4007 }
4008
4009 Node* SuperWord::last_node(Node* nd) {
4010 for (int ii = 0; ii < _iteration_last.length(); ii++) {
4011 Node* nnn = _iteration_last.at(ii);
4012 if (same_origin_idx(nnn, nd)) {
4013 #ifndef PRODUCT
4014 if (_vector_loop_debug) {
4015 tty->print_cr("SuperWord::last_node _clone_map.idx(nnn->_idx)=%d, _clone_map.idx(nd->_idx)=%d",
4016 _clone_map.idx(nnn->_idx), _clone_map.idx(nd->_idx));
4017 }
4018 #endif
4019 return nnn;
4020 }
4021 }
4022 return 0;
4023 }
4024
4025 int SuperWord::mark_generations() {
4026 Node *ii_err = NULL, *tail_err = NULL;
4027 for (int i = 0; i < _mem_slice_head.length(); i++) {
4028 Node* phi = _mem_slice_head.at(i);
4029 assert(phi->is_Phi(), "must be phi");
4030
4031 Node* tail = _mem_slice_tail.at(i);
4032 if (_ii_last == -1) {
4033 tail_err = tail;
4034 _ii_last = _clone_map.gen(tail->_idx);
4035 }
4036 else if (_ii_last != _clone_map.gen(tail->_idx)) {
4037 #ifndef PRODUCT
4038 if (TraceSuperWord && Verbose) {
4039 tty->print_cr("SuperWord::mark_generations _ii_last error - found different generations in two tail nodes ");
4040 tail->dump();
4041 tail_err->dump();
4042 }
4043 #endif
4044 return -1;
4045 }
4046
4047 // find first iteration in the loop
4048 for (DUIterator_Fast imax, i = phi->fast_outs(imax); i < imax; i++) {
4049 Node* ii = phi->fast_out(i);
4050 if (in_bb(ii) && ii->is_Store()) { // we speculate that normally Stores of one and one only generation have deps from mem phi
4051 if (_ii_first == -1) {
4052 ii_err = ii;
4053 _ii_first = _clone_map.gen(ii->_idx);
4054 } else if (_ii_first != _clone_map.gen(ii->_idx)) {
4055 #ifndef PRODUCT
4056 if (TraceSuperWord && Verbose) {
4057 tty->print_cr("SuperWord::mark_generations: _ii_first was found before and not equal to one in this node (%d)", _ii_first);
4058 ii->dump();
4059 if (ii_err!= 0) {
4060 ii_err->dump();
4061 }
4062 }
4063 #endif
4064 return -1; // this phi has Stores from different generations of unroll and cannot be simd/vectorized
4065 }
4066 }
4067 }//for (DUIterator_Fast imax,
4068 }//for (int i...
4069
4070 if (_ii_first == -1 || _ii_last == -1) {
4071 if (TraceSuperWord && Verbose) {
4072 tty->print_cr("SuperWord::mark_generations unknown error, something vent wrong");
4073 }
4074 return -1; // something vent wrong
4075 }
4076 // collect nodes in the first and last generations
4077 assert(_iteration_first.length() == 0, "_iteration_first must be empty");
4078 assert(_iteration_last.length() == 0, "_iteration_last must be empty");
4079 for (int j = 0; j < _block.length(); j++) {
4080 Node* n = _block.at(j);
4081 node_idx_t gen = _clone_map.gen(n->_idx);
4082 if ((signed)gen == _ii_first) {
4083 _iteration_first.push(n);
4084 } else if ((signed)gen == _ii_last) {
4085 _iteration_last.push(n);
4086 }
4087 }
4088
4089 // building order of iterations
4090 if (_ii_order.length() == 0 && ii_err != 0) {
4091 assert(in_bb(ii_err) && ii_err->is_Store(), "should be Store in bb");
4092 Node* nd = ii_err;
4093 while(_clone_map.gen(nd->_idx) != _ii_last) {
4094 _ii_order.push(_clone_map.gen(nd->_idx));
4095 bool found = false;
4096 for (DUIterator_Fast imax, i = nd->fast_outs(imax); i < imax; i++) {
4097 Node* use = nd->fast_out(i);
4098 if (same_origin_idx(use, nd) && use->as_Store()->in(MemNode::Memory) == nd) {
4099 found = true;
4100 nd = use;
4101 break;
4102 }
4103 }//for
4104
4105 if (found == false) {
4106 if (TraceSuperWord && Verbose) {
4107 tty->print_cr("SuperWord::mark_generations: Cannot build order of iterations - no dependent Store for %d", nd->_idx);
4108 }
4109 _ii_order.clear();
4110 return -1;
4111 }
4112 } //while
4113 _ii_order.push(_clone_map.gen(nd->_idx));
4114 }
4115
4116 #ifndef PRODUCT
4117 if (_vector_loop_debug) {
4118 tty->print_cr("SuperWord::mark_generations");
4119 tty->print_cr("First generation (%d) nodes:", _ii_first);
4120 for (int ii = 0; ii < _iteration_first.length(); ii++) _iteration_first.at(ii)->dump();
4121 tty->print_cr("Last generation (%d) nodes:", _ii_last);
4122 for (int ii = 0; ii < _iteration_last.length(); ii++) _iteration_last.at(ii)->dump();
4123 tty->print_cr(" ");
4124
4125 tty->print("SuperWord::List of generations: ");
4126 for (int jj = 0; jj < _ii_order.length(); ++jj) {
4127 tty->print("%d:%d ", jj, _ii_order.at(jj));
4128 }
4129 tty->print_cr(" ");
4130 }
4131 #endif
4132
4133 return _ii_first;
4134 }
4135
4136 bool SuperWord::fix_commutative_inputs(Node* gold, Node* fix) {
4137 assert(gold->is_Add() && fix->is_Add() || gold->is_Mul() && fix->is_Mul(), "should be only Add or Mul nodes");
4138 assert(same_origin_idx(gold, fix), "should be clones of the same node");
4139 Node* gin1 = gold->in(1);
4140 Node* gin2 = gold->in(2);
4141 Node* fin1 = fix->in(1);
4142 Node* fin2 = fix->in(2);
4143 bool swapped = false;
4144
4145 if (in_bb(gin1) && in_bb(gin2) && in_bb(fin1) && in_bb(fin1)) {
4146 if (same_origin_idx(gin1, fin1) &&
4147 same_origin_idx(gin2, fin2)) {
4148 return true; // nothing to fix
4149 }
4150 if (same_origin_idx(gin1, fin2) &&
4151 same_origin_idx(gin2, fin1)) {
4152 fix->swap_edges(1, 2);
4153 swapped = true;
4154 }
4155 }
4156 // at least one input comes from outside of bb
4157 if (gin1->_idx == fin1->_idx) {
4158 return true; // nothing to fix
4159 }
4160 if (!swapped && (gin1->_idx == fin2->_idx || gin2->_idx == fin1->_idx)) { //swapping is expensive, check condition first
4161 fix->swap_edges(1, 2);
4162 swapped = true;
4163 }
4164
4165 if (swapped) {
4166 #ifndef PRODUCT
4167 if (_vector_loop_debug) {
4168 tty->print_cr("SuperWord::fix_commutative_inputs: fixed node %d", fix->_idx);
4169 }
4170 #endif
4171 return true;
4172 }
4173
4174 if (TraceSuperWord && Verbose) {
4175 tty->print_cr("SuperWord::fix_commutative_inputs: cannot fix node %d", fix->_idx);
4176 }
4177
4178 return false;
4179 }
4180
4181 bool SuperWord::pack_parallel() {
4182 #ifndef PRODUCT
4183 if (_vector_loop_debug) {
4184 tty->print_cr("SuperWord::pack_parallel: START");
4185 }
4186 #endif
4187
4188 _packset.clear();
4189
4190 for (int ii = 0; ii < _iteration_first.length(); ii++) {
4191 Node* nd = _iteration_first.at(ii);
4192 if (in_bb(nd) && (nd->is_Load() || nd->is_Store() || nd->is_Add() || nd->is_Mul())) {
4193 Node_List* pk = new Node_List();
4194 pk->push(nd);
4195 for (int gen = 1; gen < _ii_order.length(); ++gen) {
4196 for (int kk = 0; kk < _block.length(); kk++) {
4197 Node* clone = _block.at(kk);
4198 if (same_origin_idx(clone, nd) &&
4199 _clone_map.gen(clone->_idx) == _ii_order.at(gen)) {
4200 if (nd->is_Add() || nd->is_Mul()) {
4201 fix_commutative_inputs(nd, clone);
4202 }
4203 pk->push(clone);
4204 if (pk->size() == 4) {
4205 _packset.append(pk);
4206 #ifndef PRODUCT
4207 if (_vector_loop_debug) {
4208 tty->print_cr("SuperWord::pack_parallel: added pack ");
4209 pk->dump();
4210 }
4211 #endif
4212 if (_clone_map.gen(clone->_idx) != _ii_last) {
4213 pk = new Node_List();
4214 }
4215 }
4216 break;
4217 }
4218 }
4219 }//for
4220 }//if
4221 }//for
4222
4223 #ifndef PRODUCT
4224 if (_vector_loop_debug) {
4225 tty->print_cr("SuperWord::pack_parallel: END");
4226 }
4227 #endif
4228
4229 return true;
4230 }
4231
4232 bool SuperWord::hoist_loads_in_graph() {
4233 GrowableArray<Node*> loads;
4234
4235 #ifndef PRODUCT
4236 if (_vector_loop_debug) {
4237 tty->print_cr("SuperWord::hoist_loads_in_graph: total number _mem_slice_head.length() = %d", _mem_slice_head.length());
4238 }
4239 #endif
4240
4241 for (int i = 0; i < _mem_slice_head.length(); i++) {
4242 Node* n = _mem_slice_head.at(i);
4243 if ( !in_bb(n) || !n->is_Phi() || n->bottom_type() != Type::MEMORY) {
4244 if (TraceSuperWord && Verbose) {
4245 tty->print_cr("SuperWord::hoist_loads_in_graph: skipping unexpected node n=%d", n->_idx);
4246 }
4247 continue;
4248 }
4249
4250 #ifndef PRODUCT
4251 if (_vector_loop_debug) {
4252 tty->print_cr("SuperWord::hoist_loads_in_graph: processing phi %d = _mem_slice_head.at(%d);", n->_idx, i);
4253 }
4254 #endif
4255
4256 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4257 Node* ld = n->fast_out(i);
4258 if (ld->is_Load() && ld->as_Load()->in(MemNode::Memory) == n && in_bb(ld)) {
4259 for (int i = 0; i < _block.length(); i++) {
4260 Node* ld2 = _block.at(i);
4261 if (ld2->is_Load() && same_origin_idx(ld, ld2) &&
4262 !same_generation(ld, ld2)) { // <= do not collect the first generation ld
4263 #ifndef PRODUCT
4264 if (_vector_loop_debug) {
4265 tty->print_cr("SuperWord::hoist_loads_in_graph: will try to hoist load ld2->_idx=%d, cloned from %d (ld->_idx=%d)",
4266 ld2->_idx, _clone_map.idx(ld->_idx), ld->_idx);
4267 }
4268 #endif
4269 // could not do on-the-fly, since iterator is immutable
4270 loads.push(ld2);
4271 }
4272 }// for
4273 }//if
4274 }//for (DUIterator_Fast imax,
4275 }//for (int i = 0; i
4276
4277 for (int i = 0; i < loads.length(); i++) {
4278 LoadNode* ld = loads.at(i)->as_Load();
4279 Node* phi = find_phi_for_mem_dep(ld);
4280 if (phi != NULL) {
4281 #ifndef PRODUCT
4282 if (_vector_loop_debug) {
4283 tty->print_cr("SuperWord::hoist_loads_in_graph replacing MemNode::Memory(%d) edge in %d with one from %d",
4284 MemNode::Memory, ld->_idx, phi->_idx);
4285 }
4286 #endif
4287 _igvn.replace_input_of(ld, MemNode::Memory, phi);
4288 }
4289 }//for
4290
4291 restart(); // invalidate all basic structures, since we rebuilt the graph
4292
4293 if (TraceSuperWord && Verbose) {
4294 tty->print_cr("\nSuperWord::hoist_loads_in_graph() the graph was rebuilt, all structures invalidated and need rebuild");
4295 }
4296
4297 return true;
4298 }
4299