1 /*
2 * Copyright (c) 2007, 2015, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 */
23
24 #include "precompiled.hpp"
25 #include "compiler/compileLog.hpp"
26 #include "libadt/vectset.hpp"
27 #include "memory/allocation.inline.hpp"
28 #include "opto/addnode.hpp"
29 #include "opto/callnode.hpp"
30 #include "opto/castnode.hpp"
31 #include "opto/convertnode.hpp"
32 #include "opto/divnode.hpp"
33 #include "opto/matcher.hpp"
34 #include "opto/memnode.hpp"
35 #include "opto/mulnode.hpp"
36 #include "opto/opcodes.hpp"
37 #include "opto/opaquenode.hpp"
38 #include "opto/superword.hpp"
39 #include "opto/vectornode.hpp"
40 #include "opto/movenode.hpp"
41
42 //
43 // S U P E R W O R D T R A N S F O R M
44 //=============================================================================
45
46 //------------------------------SuperWord---------------------------
47 SuperWord::SuperWord(PhaseIdealLoop* phase) :
48 _phase(phase),
49 _igvn(phase->_igvn),
50 _arena(phase->C->comp_arena()),
51 _packset(arena(), 8, 0, NULL), // packs for the current block
52 _bb_idx(arena(), (int)(1.10 * phase->C->unique()), 0, 0), // node idx to index in bb
53 _block(arena(), 8, 0, NULL), // nodes in current block
54 _data_entry(arena(), 8, 0, NULL), // nodes with all inputs from outside
55 _mem_slice_head(arena(), 8, 0, NULL), // memory slice heads
56 _mem_slice_tail(arena(), 8, 0, NULL), // memory slice tails
57 _node_info(arena(), 8, 0, SWNodeInfo::initial), // info needed per node
58 _clone_map(phase->C->clone_map()), // map of nodes created in cloning
59 _cmovev_kit(_arena, this), // map to facilitate CMoveVD creation
60 _align_to_ref(NULL), // memory reference to align vectors to
61 _disjoint_ptrs(arena(), 8, 0, OrderedPair::initial), // runtime disambiguated pointer pairs
62 _dg(_arena), // dependence graph
63 _visited(arena()), // visited node set
64 _post_visited(arena()), // post visited node set
65 _n_idx_list(arena(), 8), // scratch list of (node,index) pairs
66 _stk(arena(), 8, 0, NULL), // scratch stack of nodes
67 _nlist(arena(), 8, 0, NULL), // scratch list of nodes
68 _lpt(NULL), // loop tree node
69 _lp(NULL), // LoopNode
70 _bb(NULL), // basic block
71 _iv(NULL), // induction var
72 _race_possible(false), // cases where SDMU is true
73 _early_return(true), // analysis evaluations routine
74 _num_work_vecs(0), // amount of vector work we have
75 _num_reductions(0), // amount of reduction work we have
76 _do_vector_loop(phase->C->do_vector_loop()), // whether to do vectorization/simd style
77 _do_reserve_copy(DoReserveCopyInSuperWord),
78 _ii_first(-1), // first loop generation index - only if do_vector_loop()
79 _ii_last(-1), // last loop generation index - only if do_vector_loop()
80 _ii_order(arena(), 8, 0, 0)
81 {
82 #ifndef PRODUCT
83 _vector_loop_debug = 0;
84 if (_phase->C->method() != NULL) {
85 _vector_loop_debug = phase->C->directive()->VectorizeDebugOption;
86 }
87
88 #endif
89 }
90
91 //------------------------------transform_loop---------------------------
92 void SuperWord::transform_loop(IdealLoopTree* lpt, bool do_optimization) {
93 assert(UseSuperWord, "should be");
94 // Do vectors exist on this architecture?
95 if (Matcher::vector_width_in_bytes(T_BYTE) < 2) return;
96
97 assert(lpt->_head->is_CountedLoop(), "must be");
98 CountedLoopNode *cl = lpt->_head->as_CountedLoop();
99
100 if (!cl->is_valid_counted_loop()) return; // skip malformed counted loop
101
102 if (!cl->is_main_loop() ) return; // skip normal, pre, and post loops
103 // Check for no control flow in body (other than exit)
104 Node *cl_exit = cl->loopexit();
105 if (cl_exit->in(0) != lpt->_head) {
106 #ifndef PRODUCT
107 if (TraceSuperWord) {
108 tty->print_cr("SuperWord::transform_loop: loop too complicated, cl_exit->in(0) != lpt->_head");
109 tty->print("cl_exit %d", cl_exit->_idx); cl_exit->dump();
110 tty->print("cl_exit->in(0) %d", cl_exit->in(0)->_idx); cl_exit->in(0)->dump();
111 tty->print("lpt->_head %d", lpt->_head->_idx); lpt->_head->dump();
112 lpt->dump_head();
113 }
114 #endif
115 return;
116 }
117
118 // Make sure the are no extra control users of the loop backedge
119 if (cl->back_control()->outcnt() != 1) {
120 return;
121 }
122
123 // We only re-enter slp when we vector mapped a queried loop and we want to
124 // continue unrolling, in this case, slp is not subsequently done.
125 if (cl->do_unroll_only()) return;
126
127 // Check for pre-loop ending with CountedLoopEnd(Bool(Cmp(x,Opaque1(limit))))
128 CountedLoopEndNode* pre_end = get_pre_loop_end(cl);
129 if (pre_end == NULL) return;
130 Node *pre_opaq1 = pre_end->limit();
131 if (pre_opaq1->Opcode() != Op_Opaque1) return;
132
133 init(); // initialize data structures
134
135 set_lpt(lpt);
136 set_lp(cl);
137
138 // For now, define one block which is the entire loop body
139 set_bb(cl);
140
141 if (do_optimization) {
142 assert(_packset.length() == 0, "packset must be empty");
143 SLP_extract();
144 }
145 }
146
147 //------------------------------early unrolling analysis------------------------------
148 void SuperWord::unrolling_analysis(int &local_loop_unroll_factor) {
149 bool is_slp = true;
150 ResourceMark rm;
151 size_t ignored_size = lpt()->_body.size();
152 int *ignored_loop_nodes = NEW_RESOURCE_ARRAY(int, ignored_size);
153 Node_Stack nstack((int)ignored_size);
154 CountedLoopNode *cl = lpt()->_head->as_CountedLoop();
155 Node *cl_exit = cl->loopexit();
156
157 // First clear the entries
158 for (uint i = 0; i < lpt()->_body.size(); i++) {
159 ignored_loop_nodes[i] = -1;
160 }
161
162 int max_vector = Matcher::max_vector_size(T_INT);
163
164 // Process the loop, some/all of the stack entries will not be in order, ergo
165 // need to preprocess the ignored initial state before we process the loop
166 for (uint i = 0; i < lpt()->_body.size(); i++) {
167 Node* n = lpt()->_body.at(i);
168 if (n == cl->incr() ||
169 n->is_reduction() ||
170 n->is_AddP() ||
171 n->is_Cmp() ||
172 n->is_IfTrue() ||
173 n->is_CountedLoop() ||
174 (n == cl_exit)) {
175 ignored_loop_nodes[i] = n->_idx;
176 continue;
177 }
178
179 if (n->is_If()) {
180 IfNode *iff = n->as_If();
181 if (iff->_fcnt != COUNT_UNKNOWN && iff->_prob != PROB_UNKNOWN) {
182 if (lpt()->is_loop_exit(iff)) {
183 ignored_loop_nodes[i] = n->_idx;
184 continue;
185 }
186 }
187 }
188
189 if (n->is_Phi() && (n->bottom_type() == Type::MEMORY)) {
190 Node* n_tail = n->in(LoopNode::LoopBackControl);
191 if (n_tail != n->in(LoopNode::EntryControl)) {
192 if (!n_tail->is_Mem()) {
193 is_slp = false;
194 break;
195 }
196 }
197 }
198
199 // This must happen after check of phi/if
200 if (n->is_Phi() || n->is_If()) {
201 ignored_loop_nodes[i] = n->_idx;
202 continue;
203 }
204
205 if (n->is_LoadStore() || n->is_MergeMem() ||
206 (n->is_Proj() && !n->as_Proj()->is_CFG())) {
207 is_slp = false;
208 break;
209 }
210
211 // Ignore nodes with non-primitive type.
212 BasicType bt;
213 if (n->is_Mem()) {
214 bt = n->as_Mem()->memory_type();
215 } else {
216 bt = n->bottom_type()->basic_type();
217 }
218 if (is_java_primitive(bt) == false) {
219 ignored_loop_nodes[i] = n->_idx;
220 continue;
221 }
222
223 if (n->is_Mem()) {
224 MemNode* current = n->as_Mem();
225 Node* adr = n->in(MemNode::Address);
226 Node* n_ctrl = _phase->get_ctrl(adr);
227
228 // save a queue of post process nodes
229 if (n_ctrl != NULL && lpt()->is_member(_phase->get_loop(n_ctrl))) {
230 // Process the memory expression
231 int stack_idx = 0;
232 bool have_side_effects = true;
233 if (adr->is_AddP() == false) {
234 nstack.push(adr, stack_idx++);
235 } else {
236 // Mark the components of the memory operation in nstack
237 SWPointer p1(current, this, &nstack, true);
238 have_side_effects = p1.node_stack()->is_nonempty();
239 }
240
241 // Process the pointer stack
242 while (have_side_effects) {
243 Node* pointer_node = nstack.node();
244 for (uint j = 0; j < lpt()->_body.size(); j++) {
245 Node* cur_node = lpt()->_body.at(j);
246 if (cur_node == pointer_node) {
247 ignored_loop_nodes[j] = cur_node->_idx;
248 break;
249 }
250 }
251 nstack.pop();
252 have_side_effects = nstack.is_nonempty();
253 }
254 }
255 }
256 }
257
258 if (is_slp) {
259 // Now we try to find the maximum supported consistent vector which the machine
260 // description can use
261 for (uint i = 0; i < lpt()->_body.size(); i++) {
262 if (ignored_loop_nodes[i] != -1) continue;
263
264 BasicType bt;
265 Node* n = lpt()->_body.at(i);
266 if (n->is_Mem()) {
267 bt = n->as_Mem()->memory_type();
268 } else {
269 bt = n->bottom_type()->basic_type();
270 }
271 if (is_java_primitive(bt) == false) continue;
272
273 int cur_max_vector = Matcher::max_vector_size(bt);
274
275 // If a max vector exists which is not larger than _local_loop_unroll_factor
276 // stop looking, we already have the max vector to map to.
277 if (cur_max_vector < local_loop_unroll_factor) {
278 is_slp = false;
279 NOT_PRODUCT(if (TraceSuperWordLoopUnrollAnalysis) tty->print_cr("slp analysis fails: unroll limit greater than max vector\n"));
280 break;
281 }
282
283 // Map the maximal common vector
284 if (VectorNode::implemented(n->Opcode(), cur_max_vector, bt)) {
285 if (cur_max_vector < max_vector) {
286 max_vector = cur_max_vector;
287 }
288 }
289 }
290 if (is_slp) {
291 local_loop_unroll_factor = max_vector;
292 cl->mark_passed_slp();
293 }
294 cl->mark_was_slp();
295 cl->set_slp_max_unroll(local_loop_unroll_factor);
296 }
297 }
298
299 //------------------------------SLP_extract---------------------------
300 // Extract the superword level parallelism
301 //
302 // 1) A reverse post-order of nodes in the block is constructed. By scanning
303 // this list from first to last, all definitions are visited before their uses.
304 //
305 // 2) A point-to-point dependence graph is constructed between memory references.
306 // This simplies the upcoming "independence" checker.
307 //
308 // 3) The maximum depth in the node graph from the beginning of the block
309 // to each node is computed. This is used to prune the graph search
310 // in the independence checker.
311 //
312 // 4) For integer types, the necessary bit width is propagated backwards
313 // from stores to allow packed operations on byte, char, and short
314 // integers. This reverses the promotion to type "int" that javac
315 // did for operations like: char c1,c2,c3; c1 = c2 + c3.
316 //
317 // 5) One of the memory references is picked to be an aligned vector reference.
318 // The pre-loop trip count is adjusted to align this reference in the
319 // unrolled body.
320 //
321 // 6) The initial set of pack pairs is seeded with memory references.
322 //
323 // 7) The set of pack pairs is extended by following use->def and def->use links.
324 //
325 // 8) The pairs are combined into vector sized packs.
326 //
327 // 9) Reorder the memory slices to co-locate members of the memory packs.
328 //
329 // 10) Generate ideal vector nodes for the final set of packs and where necessary,
330 // inserting scalar promotion, vector creation from multiple scalars, and
331 // extraction of scalar values from vectors.
332 //
333 void SuperWord::SLP_extract() {
334
335 #ifndef PRODUCT
336 if (_do_vector_loop && TraceSuperWord) {
337 tty->print("SuperWord::SLP_extract\n");
338 tty->print("input loop\n");
339 _lpt->dump_head();
340 _lpt->dump();
341 for (uint i = 0; i < _lpt->_body.size(); i++) {
342 _lpt->_body.at(i)->dump();
343 }
344 }
345 #endif
346 // Ready the block
347 if (!construct_bb()) {
348 return; // Exit if no interesting nodes or complex graph.
349 }
350 // build _dg, _disjoint_ptrs
351 dependence_graph();
352
353 // compute function depth(Node*)
354 compute_max_depth();
355
356 if (_do_vector_loop) {
357 if (mark_generations() != -1) {
358 hoist_loads_in_graph(); // this only rebuild the graph; all basic structs need rebuild explicitly
359
360 if (!construct_bb()) {
361 return; // Exit if no interesting nodes or complex graph.
362 }
363 dependence_graph();
364 compute_max_depth();
365 }
366
367 #ifndef PRODUCT
368 if (TraceSuperWord) {
369 tty->print_cr("\nSuperWord::_do_vector_loop: graph after hoist_loads_in_graph");
370 _lpt->dump_head();
371 for (int j = 0; j < _block.length(); j++) {
372 Node* n = _block.at(j);
373 int d = depth(n);
374 for (int i = 0; i < d; i++) tty->print("%s", " ");
375 tty->print("%d :", d);
376 n->dump();
377 }
378 }
379 #endif
380 }
381
382 compute_vector_element_type();
383
384 // Attempt vectorization
385
386 find_adjacent_refs();
387
388 extend_packlist();
389
390 if (_do_vector_loop) {
391 if (_packset.length() == 0) {
392 #ifndef PRODUCT
393 if (TraceSuperWord) {
394 tty->print_cr("\nSuperWord::_do_vector_loop DFA could not build packset, now trying to build anyway");
395 }
396 #endif
397 pack_parallel();
398 }
399 }
400
401 combine_packs();
402
403 construct_my_pack_map();
404
405 if (_do_vector_loop) {
406 merge_packs_to_cmovd();
407 }
408
409 filter_packs();
410
411 schedule();
412
413 output();
414 }
415
416 //------------------------------find_adjacent_refs---------------------------
417 // Find the adjacent memory references and create pack pairs for them.
418 // This is the initial set of packs that will then be extended by
419 // following use->def and def->use links. The align positions are
420 // assigned relative to the reference "align_to_ref"
421 void SuperWord::find_adjacent_refs() {
422 // Get list of memory operations
423 Node_List memops;
424 for (int i = 0; i < _block.length(); i++) {
425 Node* n = _block.at(i);
426 if (n->is_Mem() && !n->is_LoadStore() && in_bb(n) &&
427 is_java_primitive(n->as_Mem()->memory_type())) {
428 int align = memory_alignment(n->as_Mem(), 0);
429 if (align != bottom_align) {
430 memops.push(n);
431 }
432 }
433 }
434
435 Node_List align_to_refs;
436 int best_iv_adjustment = 0;
437 MemNode* best_align_to_mem_ref = NULL;
438
439 while (memops.size() != 0) {
440 // Find a memory reference to align to.
441 MemNode* mem_ref = find_align_to_ref(memops);
442 if (mem_ref == NULL) break;
443 align_to_refs.push(mem_ref);
444 int iv_adjustment = get_iv_adjustment(mem_ref);
445
446 if (best_align_to_mem_ref == NULL) {
447 // Set memory reference which is the best from all memory operations
448 // to be used for alignment. The pre-loop trip count is modified to align
449 // this reference to a vector-aligned address.
450 best_align_to_mem_ref = mem_ref;
451 best_iv_adjustment = iv_adjustment;
452 NOT_PRODUCT(find_adjacent_refs_trace_1(best_align_to_mem_ref, best_iv_adjustment);)
453 }
454
455 SWPointer align_to_ref_p(mem_ref, this, NULL, false);
456 // Set alignment relative to "align_to_ref" for all related memory operations.
457 for (int i = memops.size() - 1; i >= 0; i--) {
458 MemNode* s = memops.at(i)->as_Mem();
459 if (isomorphic(s, mem_ref) &&
460 (!_do_vector_loop || same_origin_idx(s, mem_ref))) {
461 SWPointer p2(s, this, NULL, false);
462 if (p2.comparable(align_to_ref_p)) {
463 int align = memory_alignment(s, iv_adjustment);
464 set_alignment(s, align);
465 }
466 }
467 }
468
469 // Create initial pack pairs of memory operations for which
470 // alignment is set and vectors will be aligned.
471 bool create_pack = true;
472 if (memory_alignment(mem_ref, best_iv_adjustment) == 0 || _do_vector_loop) {
473 if (!Matcher::misaligned_vectors_ok()) {
474 int vw = vector_width(mem_ref);
475 int vw_best = vector_width(best_align_to_mem_ref);
476 if (vw > vw_best) {
477 // Do not vectorize a memory access with more elements per vector
478 // if unaligned memory access is not allowed because number of
479 // iterations in pre-loop will be not enough to align it.
480 create_pack = false;
481 } else {
482 SWPointer p2(best_align_to_mem_ref, this, NULL, false);
483 if (align_to_ref_p.invar() != p2.invar()) {
484 // Do not vectorize memory accesses with different invariants
485 // if unaligned memory accesses are not allowed.
486 create_pack = false;
487 }
488 }
489 }
490 } else {
491 if (same_velt_type(mem_ref, best_align_to_mem_ref)) {
492 // Can't allow vectorization of unaligned memory accesses with the
493 // same type since it could be overlapped accesses to the same array.
494 create_pack = false;
495 } else {
496 // Allow independent (different type) unaligned memory operations
497 // if HW supports them.
498 if (!Matcher::misaligned_vectors_ok()) {
499 create_pack = false;
500 } else {
501 // Check if packs of the same memory type but
502 // with a different alignment were created before.
503 for (uint i = 0; i < align_to_refs.size(); i++) {
504 MemNode* mr = align_to_refs.at(i)->as_Mem();
505 if (same_velt_type(mr, mem_ref) &&
506 memory_alignment(mr, iv_adjustment) != 0)
507 create_pack = false;
508 }
509 }
510 }
511 }
512 if (create_pack) {
513 for (uint i = 0; i < memops.size(); i++) {
514 Node* s1 = memops.at(i);
515 int align = alignment(s1);
516 if (align == top_align) continue;
517 for (uint j = 0; j < memops.size(); j++) {
518 Node* s2 = memops.at(j);
519 if (alignment(s2) == top_align) continue;
520 if (s1 != s2 && are_adjacent_refs(s1, s2)) {
521 if (stmts_can_pack(s1, s2, align)) {
522 Node_List* pair = new Node_List();
523 pair->push(s1);
524 pair->push(s2);
525 if (!_do_vector_loop || same_origin_idx(s1, s2)) {
526 _packset.append(pair);
527 }
528 }
529 }
530 }
531 }
532 } else { // Don't create unaligned pack
533 // First, remove remaining memory ops of the same type from the list.
534 for (int i = memops.size() - 1; i >= 0; i--) {
535 MemNode* s = memops.at(i)->as_Mem();
536 if (same_velt_type(s, mem_ref)) {
537 memops.remove(i);
538 }
539 }
540
541 // Second, remove already constructed packs of the same type.
542 for (int i = _packset.length() - 1; i >= 0; i--) {
543 Node_List* p = _packset.at(i);
544 MemNode* s = p->at(0)->as_Mem();
545 if (same_velt_type(s, mem_ref)) {
546 remove_pack_at(i);
547 }
548 }
549
550 // If needed find the best memory reference for loop alignment again.
551 if (same_velt_type(mem_ref, best_align_to_mem_ref)) {
552 // Put memory ops from remaining packs back on memops list for
553 // the best alignment search.
554 uint orig_msize = memops.size();
555 for (int i = 0; i < _packset.length(); i++) {
556 Node_List* p = _packset.at(i);
557 MemNode* s = p->at(0)->as_Mem();
558 assert(!same_velt_type(s, mem_ref), "sanity");
559 memops.push(s);
560 }
561 MemNode* best_align_to_mem_ref = find_align_to_ref(memops);
562 if (best_align_to_mem_ref == NULL) {
563 NOT_PRODUCT(if (TraceSuperWord) tty->print_cr("SuperWord::find_adjacent_refs(): best_align_to_mem_ref == NULL");)
564 break;
565 }
566 best_iv_adjustment = get_iv_adjustment(best_align_to_mem_ref);
567 NOT_PRODUCT(find_adjacent_refs_trace_1(best_align_to_mem_ref, best_iv_adjustment);)
568 // Restore list.
569 while (memops.size() > orig_msize)
570 (void)memops.pop();
571 }
572 } // unaligned memory accesses
573
574 // Remove used mem nodes.
575 for (int i = memops.size() - 1; i >= 0; i--) {
576 MemNode* m = memops.at(i)->as_Mem();
577 if (alignment(m) != top_align) {
578 memops.remove(i);
579 }
580 }
581
582 } // while (memops.size() != 0
583 set_align_to_ref(best_align_to_mem_ref);
584
585 #ifndef PRODUCT
586 if (TraceSuperWord) {
587 tty->print_cr("\nAfter find_adjacent_refs");
588 print_packset();
589 }
590 #endif
591 }
592
593 #ifndef PRODUCT
594 void SuperWord::find_adjacent_refs_trace_1(Node* best_align_to_mem_ref, int best_iv_adjustment) {
595 if (is_trace_adjacent()) {
596 tty->print("SuperWord::find_adjacent_refs best_align_to_mem_ref = %d, best_iv_adjustment = %d",
597 best_align_to_mem_ref->_idx, best_iv_adjustment);
598 best_align_to_mem_ref->dump();
599 }
600 }
601 #endif
602
603 //------------------------------find_align_to_ref---------------------------
604 // Find a memory reference to align the loop induction variable to.
605 // Looks first at stores then at loads, looking for a memory reference
606 // with the largest number of references similar to it.
607 MemNode* SuperWord::find_align_to_ref(Node_List &memops) {
608 GrowableArray<int> cmp_ct(arena(), memops.size(), memops.size(), 0);
609
610 // Count number of comparable memory ops
611 for (uint i = 0; i < memops.size(); i++) {
612 MemNode* s1 = memops.at(i)->as_Mem();
613 SWPointer p1(s1, this, NULL, false);
614 // Discard if pre loop can't align this reference
615 if (!ref_is_alignable(p1)) {
616 *cmp_ct.adr_at(i) = 0;
617 continue;
618 }
619 for (uint j = i+1; j < memops.size(); j++) {
620 MemNode* s2 = memops.at(j)->as_Mem();
621 if (isomorphic(s1, s2)) {
622 SWPointer p2(s2, this, NULL, false);
623 if (p1.comparable(p2)) {
624 (*cmp_ct.adr_at(i))++;
625 (*cmp_ct.adr_at(j))++;
626 }
627 }
628 }
629 }
630
631 // Find Store (or Load) with the greatest number of "comparable" references,
632 // biggest vector size, smallest data size and smallest iv offset.
633 int max_ct = 0;
634 int max_vw = 0;
635 int max_idx = -1;
636 int min_size = max_jint;
637 int min_iv_offset = max_jint;
638 for (uint j = 0; j < memops.size(); j++) {
639 MemNode* s = memops.at(j)->as_Mem();
640 if (s->is_Store()) {
641 int vw = vector_width_in_bytes(s);
642 assert(vw > 1, "sanity");
643 SWPointer p(s, this, NULL, false);
644 if (cmp_ct.at(j) > max_ct ||
645 cmp_ct.at(j) == max_ct &&
646 (vw > max_vw ||
647 vw == max_vw &&
648 (data_size(s) < min_size ||
649 data_size(s) == min_size &&
650 (p.offset_in_bytes() < min_iv_offset)))) {
651 max_ct = cmp_ct.at(j);
652 max_vw = vw;
653 max_idx = j;
654 min_size = data_size(s);
655 min_iv_offset = p.offset_in_bytes();
656 }
657 }
658 }
659 // If no stores, look at loads
660 if (max_ct == 0) {
661 for (uint j = 0; j < memops.size(); j++) {
662 MemNode* s = memops.at(j)->as_Mem();
663 if (s->is_Load()) {
664 int vw = vector_width_in_bytes(s);
665 assert(vw > 1, "sanity");
666 SWPointer p(s, this, NULL, false);
667 if (cmp_ct.at(j) > max_ct ||
668 cmp_ct.at(j) == max_ct &&
669 (vw > max_vw ||
670 vw == max_vw &&
671 (data_size(s) < min_size ||
672 data_size(s) == min_size &&
673 (p.offset_in_bytes() < min_iv_offset)))) {
674 max_ct = cmp_ct.at(j);
675 max_vw = vw;
676 max_idx = j;
677 min_size = data_size(s);
678 min_iv_offset = p.offset_in_bytes();
679 }
680 }
681 }
682 }
683
684 #ifdef ASSERT
685 if (TraceSuperWord && Verbose) {
686 tty->print_cr("\nVector memops after find_align_to_ref");
687 for (uint i = 0; i < memops.size(); i++) {
688 MemNode* s = memops.at(i)->as_Mem();
689 s->dump();
690 }
691 }
692 #endif
693
694 if (max_ct > 0) {
695 #ifdef ASSERT
696 if (TraceSuperWord) {
697 tty->print("\nVector align to node: ");
698 memops.at(max_idx)->as_Mem()->dump();
699 }
700 #endif
701 return memops.at(max_idx)->as_Mem();
702 }
703 return NULL;
704 }
705
706 //------------------------------ref_is_alignable---------------------------
707 // Can the preloop align the reference to position zero in the vector?
708 bool SuperWord::ref_is_alignable(SWPointer& p) {
709 if (!p.has_iv()) {
710 return true; // no induction variable
711 }
712 CountedLoopEndNode* pre_end = get_pre_loop_end(lp()->as_CountedLoop());
713 assert(pre_end != NULL, "we must have a correct pre-loop");
714 assert(pre_end->stride_is_con(), "pre loop stride is constant");
715 int preloop_stride = pre_end->stride_con();
716
717 int span = preloop_stride * p.scale_in_bytes();
718 int mem_size = p.memory_size();
719 int offset = p.offset_in_bytes();
720 // Stride one accesses are alignable if offset is aligned to memory operation size.
721 // Offset can be unaligned when UseUnalignedAccesses is used.
722 if (ABS(span) == mem_size && (ABS(offset) % mem_size) == 0) {
723 return true;
724 }
725 // If the initial offset from start of the object is computable,
726 // check if the pre-loop can align the final offset accordingly.
727 //
728 // In other words: Can we find an i such that the offset
729 // after i pre-loop iterations is aligned to vw?
730 // (init_offset + pre_loop) % vw == 0 (1)
731 // where
732 // pre_loop = i * span
733 // is the number of bytes added to the offset by i pre-loop iterations.
734 //
735 // For this to hold we need pre_loop to increase init_offset by
736 // pre_loop = vw - (init_offset % vw)
737 //
738 // This is only possible if pre_loop is divisible by span because each
739 // pre-loop iteration increases the initial offset by 'span' bytes:
740 // (vw - (init_offset % vw)) % span == 0
741 //
742 int vw = vector_width_in_bytes(p.mem());
743 assert(vw > 1, "sanity");
744 Node* init_nd = pre_end->init_trip();
745 if (init_nd->is_Con() && p.invar() == NULL) {
746 int init = init_nd->bottom_type()->is_int()->get_con();
747 int init_offset = init * p.scale_in_bytes() + offset;
748 assert(init_offset >= 0, "positive offset from object start");
749 if (vw % span == 0) {
750 // If vm is a multiple of span, we use formula (1).
751 if (span > 0) {
752 return (vw - (init_offset % vw)) % span == 0;
753 } else {
754 assert(span < 0, "nonzero stride * scale");
755 return (init_offset % vw) % -span == 0;
756 }
757 } else if (span % vw == 0) {
758 // If span is a multiple of vw, we can simplify formula (1) to:
759 // (init_offset + i * span) % vw == 0
760 // =>
761 // (init_offset % vw) + ((i * span) % vw) == 0
762 // =>
763 // init_offset % vw == 0
764 //
765 // Because we add a multiple of vw to the initial offset, the final
766 // offset is a multiple of vw if and only if init_offset is a multiple.
767 //
768 return (init_offset % vw) == 0;
769 }
770 }
771 return false;
772 }
773
774 //---------------------------get_iv_adjustment---------------------------
775 // Calculate loop's iv adjustment for this memory ops.
776 int SuperWord::get_iv_adjustment(MemNode* mem_ref) {
777 SWPointer align_to_ref_p(mem_ref, this, NULL, false);
778 int offset = align_to_ref_p.offset_in_bytes();
779 int scale = align_to_ref_p.scale_in_bytes();
780 int elt_size = align_to_ref_p.memory_size();
781 int vw = vector_width_in_bytes(mem_ref);
782 assert(vw > 1, "sanity");
783 int iv_adjustment;
784 if (scale != 0) {
785 int stride_sign = (scale * iv_stride()) > 0 ? 1 : -1;
786 // At least one iteration is executed in pre-loop by default. As result
787 // several iterations are needed to align memory operations in main-loop even
788 // if offset is 0.
789 int iv_adjustment_in_bytes = (stride_sign * vw - (offset % vw));
790 assert(((ABS(iv_adjustment_in_bytes) % elt_size) == 0),
791 "(%d) should be divisible by (%d)", iv_adjustment_in_bytes, elt_size);
792 iv_adjustment = iv_adjustment_in_bytes/elt_size;
793 } else {
794 // This memory op is not dependent on iv (scale == 0)
795 iv_adjustment = 0;
796 }
797
798 #ifndef PRODUCT
799 if (TraceSuperWord) {
800 tty->print("SuperWord::get_iv_adjustment: n = %d, noffset = %d iv_adjust = %d elt_size = %d scale = %d iv_stride = %d vect_size %d: ",
801 mem_ref->_idx, offset, iv_adjustment, elt_size, scale, iv_stride(), vw);
802 mem_ref->dump();
803 }
804 #endif
805 return iv_adjustment;
806 }
807
808 //---------------------------dependence_graph---------------------------
809 // Construct dependency graph.
810 // Add dependence edges to load/store nodes for memory dependence
811 // A.out()->DependNode.in(1) and DependNode.out()->B.prec(x)
812 void SuperWord::dependence_graph() {
813 // First, assign a dependence node to each memory node
814 for (int i = 0; i < _block.length(); i++ ) {
815 Node *n = _block.at(i);
816 if (n->is_Mem() || n->is_Phi() && n->bottom_type() == Type::MEMORY) {
817 _dg.make_node(n);
818 }
819 }
820
821 // For each memory slice, create the dependences
822 for (int i = 0; i < _mem_slice_head.length(); i++) {
823 Node* n = _mem_slice_head.at(i);
824 Node* n_tail = _mem_slice_tail.at(i);
825
826 // Get slice in predecessor order (last is first)
827 mem_slice_preds(n_tail, n, _nlist);
828
829 #ifndef PRODUCT
830 if(TraceSuperWord && Verbose) {
831 tty->print_cr("SuperWord::dependence_graph: built a new mem slice");
832 for (int j = _nlist.length() - 1; j >= 0 ; j--) {
833 _nlist.at(j)->dump();
834 }
835 }
836 #endif
837 // Make the slice dependent on the root
838 DepMem* slice = _dg.dep(n);
839 _dg.make_edge(_dg.root(), slice);
840
841 // Create a sink for the slice
842 DepMem* slice_sink = _dg.make_node(NULL);
843 _dg.make_edge(slice_sink, _dg.tail());
844
845 // Now visit each pair of memory ops, creating the edges
846 for (int j = _nlist.length() - 1; j >= 0 ; j--) {
847 Node* s1 = _nlist.at(j);
848
849 // If no dependency yet, use slice
850 if (_dg.dep(s1)->in_cnt() == 0) {
851 _dg.make_edge(slice, s1);
852 }
853 SWPointer p1(s1->as_Mem(), this, NULL, false);
854 bool sink_dependent = true;
855 for (int k = j - 1; k >= 0; k--) {
856 Node* s2 = _nlist.at(k);
857 if (s1->is_Load() && s2->is_Load())
858 continue;
859 SWPointer p2(s2->as_Mem(), this, NULL, false);
860
861 int cmp = p1.cmp(p2);
862 if (SuperWordRTDepCheck &&
863 p1.base() != p2.base() && p1.valid() && p2.valid()) {
864 // Create a runtime check to disambiguate
865 OrderedPair pp(p1.base(), p2.base());
866 _disjoint_ptrs.append_if_missing(pp);
867 } else if (!SWPointer::not_equal(cmp)) {
868 // Possibly same address
869 _dg.make_edge(s1, s2);
870 sink_dependent = false;
871 }
872 }
873 if (sink_dependent) {
874 _dg.make_edge(s1, slice_sink);
875 }
876 }
877 #ifndef PRODUCT
878 if (TraceSuperWord) {
879 tty->print_cr("\nDependence graph for slice: %d", n->_idx);
880 for (int q = 0; q < _nlist.length(); q++) {
881 _dg.print(_nlist.at(q));
882 }
883 tty->cr();
884 }
885 #endif
886 _nlist.clear();
887 }
888
889 #ifndef PRODUCT
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 #endif
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 NOT_PRODUCT(if (TraceSuperWord && Verbose) tty->print_cr("SuperWord::mem_slice_preds: added pred(%d)", out->_idx);)
916 }
917 } else {
918 // FIXME
919 if (out->is_MergeMem() && !in_bb(out)) {
920 // Either unrolling is causing a memory edge not to disappear,
921 // or need to run igvn.optimize() again before SLP
922 } else if (out->is_Phi() && out->bottom_type() == Type::MEMORY && !in_bb(out)) {
923 // Ditto. Not sure what else to check further.
924 } else if (out->Opcode() == Op_StoreCM && out->in(MemNode::OopStore) == n) {
925 // StoreCM has an input edge used as a precedence edge.
926 // Maybe an issue when oop stores are vectorized.
927 } else {
928 assert(out == prev || prev == NULL, "no branches off of store slice");
929 }
930 }//else
931 }//for
932 if (n == stop) break;
933 preds.push(n);
934 NOT_PRODUCT(if (TraceSuperWord && Verbose) tty->print_cr("SuperWord::mem_slice_preds: added pred(%d)", n->_idx);)
935 prev = n;
936 assert(n->is_Mem(), "unexpected node %s", n->Name());
937 n = n->in(MemNode::Memory);
938 }
939 }
940
941 //------------------------------stmts_can_pack---------------------------
942 // Can s1 and s2 be in a pack with s1 immediately preceding s2 and
943 // s1 aligned at "align"
944 bool SuperWord::stmts_can_pack(Node* s1, Node* s2, int align) {
945
946 // Do not use superword for non-primitives
947 BasicType bt1 = velt_basic_type(s1);
948 BasicType bt2 = velt_basic_type(s2);
949 if(!is_java_primitive(bt1) || !is_java_primitive(bt2))
950 return false;
951 if (Matcher::max_vector_size(bt1) < 2) {
952 return false; // No vectors for this type
953 }
954
955 if (isomorphic(s1, s2)) {
956 if (independent(s1, s2) || reduction(s1, s2)) {
957 if (!exists_at(s1, 0) && !exists_at(s2, 1)) {
958 if (!s1->is_Mem() || are_adjacent_refs(s1, s2)) {
959 int s1_align = alignment(s1);
960 int s2_align = alignment(s2);
961 if (s1_align == top_align || s1_align == align) {
962 if (s2_align == top_align || s2_align == align + data_size(s1)) {
963 return true;
964 }
965 }
966 }
967 }
968 }
969 }
970 return false;
971 }
972
973 //------------------------------exists_at---------------------------
974 // Does s exist in a pack at position pos?
975 bool SuperWord::exists_at(Node* s, uint pos) {
976 for (int i = 0; i < _packset.length(); i++) {
977 Node_List* p = _packset.at(i);
978 if (p->at(pos) == s) {
979 return true;
980 }
981 }
982 return false;
983 }
984
985 //------------------------------are_adjacent_refs---------------------------
986 // Is s1 immediately before s2 in memory?
987 bool SuperWord::are_adjacent_refs(Node* s1, Node* s2) {
988 if (!s1->is_Mem() || !s2->is_Mem()) return false;
989 if (!in_bb(s1) || !in_bb(s2)) return false;
990
991 // Do not use superword for non-primitives
992 if (!is_java_primitive(s1->as_Mem()->memory_type()) ||
993 !is_java_primitive(s2->as_Mem()->memory_type())) {
994 return false;
995 }
996
997 // FIXME - co_locate_pack fails on Stores in different mem-slices, so
998 // only pack memops that are in the same alias set until that's fixed.
999 if (_phase->C->get_alias_index(s1->as_Mem()->adr_type()) !=
1000 _phase->C->get_alias_index(s2->as_Mem()->adr_type()))
1001 return false;
1002 SWPointer p1(s1->as_Mem(), this, NULL, false);
1003 SWPointer p2(s2->as_Mem(), this, NULL, false);
1004 if (p1.base() != p2.base() || !p1.comparable(p2)) return false;
1005 int diff = p2.offset_in_bytes() - p1.offset_in_bytes();
1006 return diff == data_size(s1);
1007 }
1008
1009 //------------------------------isomorphic---------------------------
1010 // Are s1 and s2 similar?
1011 bool SuperWord::isomorphic(Node* s1, Node* s2) {
1012 if (s1->Opcode() != s2->Opcode()) return false;
1013 if (s1->req() != s2->req()) return false;
1014 if (s1->in(0) != s2->in(0)) return false;
1015 if (!same_velt_type(s1, s2)) return false;
1016 return true;
1017 }
1018
1019 //------------------------------independent---------------------------
1020 // Is there no data path from s1 to s2 or s2 to s1?
1021 bool SuperWord::independent(Node* s1, Node* s2) {
1022 // assert(s1->Opcode() == s2->Opcode(), "check isomorphic first");
1023 int d1 = depth(s1);
1024 int d2 = depth(s2);
1025 if (d1 == d2) return s1 != s2;
1026 Node* deep = d1 > d2 ? s1 : s2;
1027 Node* shallow = d1 > d2 ? s2 : s1;
1028
1029 visited_clear();
1030
1031 return independent_path(shallow, deep);
1032 }
1033
1034 //------------------------------reduction---------------------------
1035 // Is there a data path between s1 and s2 and the nodes reductions?
1036 bool SuperWord::reduction(Node* s1, Node* s2) {
1037 bool retValue = false;
1038 int d1 = depth(s1);
1039 int d2 = depth(s2);
1040 if (d1 + 1 == d2) {
1041 if (s1->is_reduction() && s2->is_reduction()) {
1042 // This is an ordered set, so s1 should define s2
1043 for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
1044 Node* t1 = s1->fast_out(i);
1045 if (t1 == s2) {
1046 // both nodes are reductions and connected
1047 retValue = true;
1048 }
1049 }
1050 }
1051 }
1052
1053 return retValue;
1054 }
1055
1056 //------------------------------independent_path------------------------------
1057 // Helper for independent
1058 bool SuperWord::independent_path(Node* shallow, Node* deep, uint dp) {
1059 if (dp >= 1000) return false; // stop deep recursion
1060 visited_set(deep);
1061 int shal_depth = depth(shallow);
1062 assert(shal_depth <= depth(deep), "must be");
1063 for (DepPreds preds(deep, _dg); !preds.done(); preds.next()) {
1064 Node* pred = preds.current();
1065 if (in_bb(pred) && !visited_test(pred)) {
1066 if (shallow == pred) {
1067 return false;
1068 }
1069 if (shal_depth < depth(pred) && !independent_path(shallow, pred, dp+1)) {
1070 return false;
1071 }
1072 }
1073 }
1074 return true;
1075 }
1076
1077 //------------------------------set_alignment---------------------------
1078 void SuperWord::set_alignment(Node* s1, Node* s2, int align) {
1079 set_alignment(s1, align);
1080 if (align == top_align || align == bottom_align) {
1081 set_alignment(s2, align);
1082 } else {
1083 set_alignment(s2, align + data_size(s1));
1084 }
1085 }
1086
1087 //------------------------------data_size---------------------------
1088 int SuperWord::data_size(Node* s) {
1089 Node* use = NULL; //test if the node is a candidate for CMoveVD optimization, then return the size of CMov
1090 if (_do_vector_loop) {
1091 use = _cmovev_kit.is_Bool_candidate(s);
1092 if (use != NULL) {
1093 return data_size(use);
1094 }
1095 use = _cmovev_kit.is_CmpD_candidate(s);
1096 if (use != NULL) {
1097 return data_size(use);
1098 }
1099 }
1100 int bsize = type2aelembytes(velt_basic_type(s));
1101 assert(bsize != 0, "valid size");
1102 return bsize;
1103 }
1104
1105 //------------------------------extend_packlist---------------------------
1106 // Extend packset by following use->def and def->use links from pack members.
1107 void SuperWord::extend_packlist() {
1108 bool changed;
1109 do {
1110 packset_sort(_packset.length());
1111 changed = false;
1112 for (int i = 0; i < _packset.length(); i++) {
1113 Node_List* p = _packset.at(i);
1114 changed |= follow_use_defs(p);
1115 changed |= follow_def_uses(p);
1116 }
1117 } while (changed);
1118
1119 if (_race_possible) {
1120 for (int i = 0; i < _packset.length(); i++) {
1121 Node_List* p = _packset.at(i);
1122 order_def_uses(p);
1123 }
1124 }
1125
1126 #ifndef PRODUCT
1127 if (TraceSuperWord) {
1128 tty->print_cr("\nAfter extend_packlist");
1129 print_packset();
1130 }
1131 #endif
1132 }
1133
1134 //------------------------------follow_use_defs---------------------------
1135 // Extend the packset by visiting operand definitions of nodes in pack p
1136 bool SuperWord::follow_use_defs(Node_List* p) {
1137 assert(p->size() == 2, "just checking");
1138 Node* s1 = p->at(0);
1139 Node* s2 = p->at(1);
1140 assert(s1->req() == s2->req(), "just checking");
1141 assert(alignment(s1) + data_size(s1) == alignment(s2), "just checking");
1142
1143 if (s1->is_Load()) return false;
1144
1145 int align = alignment(s1);
1146 NOT_PRODUCT(if(is_trace_alignment()) tty->print_cr("SuperWord::follow_use_defs: s1 %d, align %d", s1->_idx, align);)
1147 bool changed = false;
1148 int start = s1->is_Store() ? MemNode::ValueIn : 1;
1149 int end = s1->is_Store() ? MemNode::ValueIn+1 : s1->req();
1150 for (int j = start; j < end; j++) {
1151 Node* t1 = s1->in(j);
1152 Node* t2 = s2->in(j);
1153 if (!in_bb(t1) || !in_bb(t2))
1154 continue;
1155 if (stmts_can_pack(t1, t2, align)) {
1156 if (est_savings(t1, t2) >= 0) {
1157 Node_List* pair = new Node_List();
1158 pair->push(t1);
1159 pair->push(t2);
1160 _packset.append(pair);
1161 NOT_PRODUCT(if(is_trace_alignment()) tty->print_cr("SuperWord::follow_use_defs: set_alignment(%d, %d, %d)", t1->_idx, t2->_idx, align);)
1162 set_alignment(t1, t2, align);
1163 changed = true;
1164 }
1165 }
1166 }
1167 return changed;
1168 }
1169
1170 //------------------------------follow_def_uses---------------------------
1171 // Extend the packset by visiting uses of nodes in pack p
1172 bool SuperWord::follow_def_uses(Node_List* p) {
1173 bool changed = false;
1174 Node* s1 = p->at(0);
1175 Node* s2 = p->at(1);
1176 assert(p->size() == 2, "just checking");
1177 assert(s1->req() == s2->req(), "just checking");
1178 assert(alignment(s1) + data_size(s1) == alignment(s2), "just checking");
1179
1180 if (s1->is_Store()) return false;
1181
1182 int align = alignment(s1);
1183 NOT_PRODUCT(if(is_trace_alignment()) tty->print_cr("SuperWord::follow_def_uses: s1 %d, align %d", s1->_idx, align);)
1184 int savings = -1;
1185 int num_s1_uses = 0;
1186 Node* u1 = NULL;
1187 Node* u2 = NULL;
1188 for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
1189 Node* t1 = s1->fast_out(i);
1190 num_s1_uses++;
1191 if (!in_bb(t1)) continue;
1192 for (DUIterator_Fast jmax, j = s2->fast_outs(jmax); j < jmax; j++) {
1193 Node* t2 = s2->fast_out(j);
1194 if (!in_bb(t2)) continue;
1195 if (!opnd_positions_match(s1, t1, s2, t2))
1196 continue;
1197 if (stmts_can_pack(t1, t2, align)) {
1198 int my_savings = est_savings(t1, t2);
1199 if (my_savings > savings) {
1200 savings = my_savings;
1201 u1 = t1;
1202 u2 = t2;
1203 }
1204 }
1205 }
1206 }
1207 if (num_s1_uses > 1) {
1208 _race_possible = true;
1209 }
1210 if (savings >= 0) {
1211 Node_List* pair = new Node_List();
1212 pair->push(u1);
1213 pair->push(u2);
1214 _packset.append(pair);
1215 NOT_PRODUCT(if(is_trace_alignment()) tty->print_cr("SuperWord::follow_def_uses: set_alignment(%d, %d, %d)", u1->_idx, u2->_idx, align);)
1216 set_alignment(u1, u2, align);
1217 changed = true;
1218 }
1219 return changed;
1220 }
1221
1222 //------------------------------order_def_uses---------------------------
1223 // For extended packsets, ordinally arrange uses packset by major component
1224 void SuperWord::order_def_uses(Node_List* p) {
1225 Node* s1 = p->at(0);
1226
1227 if (s1->is_Store()) return;
1228
1229 // reductions are always managed beforehand
1230 if (s1->is_reduction()) return;
1231
1232 for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
1233 Node* t1 = s1->fast_out(i);
1234
1235 // Only allow operand swap on commuting operations
1236 if (!t1->is_Add() && !t1->is_Mul()) {
1237 break;
1238 }
1239
1240 // Now find t1's packset
1241 Node_List* p2 = NULL;
1242 for (int j = 0; j < _packset.length(); j++) {
1243 p2 = _packset.at(j);
1244 Node* first = p2->at(0);
1245 if (t1 == first) {
1246 break;
1247 }
1248 p2 = NULL;
1249 }
1250 // Arrange all sub components by the major component
1251 if (p2 != NULL) {
1252 for (uint j = 1; j < p->size(); j++) {
1253 Node* d1 = p->at(j);
1254 Node* u1 = p2->at(j);
1255 opnd_positions_match(s1, t1, d1, u1);
1256 }
1257 }
1258 }
1259 }
1260
1261 //---------------------------opnd_positions_match-------------------------
1262 // Is the use of d1 in u1 at the same operand position as d2 in u2?
1263 bool SuperWord::opnd_positions_match(Node* d1, Node* u1, Node* d2, Node* u2) {
1264 // check reductions to see if they are marshalled to represent the reduction
1265 // operator in a specified opnd
1266 if (u1->is_reduction() && u2->is_reduction()) {
1267 // ensure reductions have phis and reduction definitions feeding the 1st operand
1268 Node* first = u1->in(2);
1269 if (first->is_Phi() || first->is_reduction()) {
1270 u1->swap_edges(1, 2);
1271 }
1272 // ensure reductions have phis and reduction definitions feeding the 1st operand
1273 first = u2->in(2);
1274 if (first->is_Phi() || first->is_reduction()) {
1275 u2->swap_edges(1, 2);
1276 }
1277 return true;
1278 }
1279
1280 uint ct = u1->req();
1281 if (ct != u2->req()) return false;
1282 uint i1 = 0;
1283 uint i2 = 0;
1284 do {
1285 for (i1++; i1 < ct; i1++) if (u1->in(i1) == d1) break;
1286 for (i2++; i2 < ct; i2++) if (u2->in(i2) == d2) break;
1287 if (i1 != i2) {
1288 if ((i1 == (3-i2)) && (u2->is_Add() || u2->is_Mul())) {
1289 // Further analysis relies on operands position matching.
1290 u2->swap_edges(i1, i2);
1291 } else {
1292 return false;
1293 }
1294 }
1295 } while (i1 < ct);
1296 return true;
1297 }
1298
1299 //------------------------------est_savings---------------------------
1300 // Estimate the savings from executing s1 and s2 as a pack
1301 int SuperWord::est_savings(Node* s1, Node* s2) {
1302 int save_in = 2 - 1; // 2 operations per instruction in packed form
1303
1304 // inputs
1305 for (uint i = 1; i < s1->req(); i++) {
1306 Node* x1 = s1->in(i);
1307 Node* x2 = s2->in(i);
1308 if (x1 != x2) {
1309 if (are_adjacent_refs(x1, x2)) {
1310 save_in += adjacent_profit(x1, x2);
1311 } else if (!in_packset(x1, x2)) {
1312 save_in -= pack_cost(2);
1313 } else {
1314 save_in += unpack_cost(2);
1315 }
1316 }
1317 }
1318
1319 // uses of result
1320 uint ct = 0;
1321 int save_use = 0;
1322 for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
1323 Node* s1_use = s1->fast_out(i);
1324 for (int j = 0; j < _packset.length(); j++) {
1325 Node_List* p = _packset.at(j);
1326 if (p->at(0) == s1_use) {
1327 for (DUIterator_Fast kmax, k = s2->fast_outs(kmax); k < kmax; k++) {
1328 Node* s2_use = s2->fast_out(k);
1329 if (p->at(p->size()-1) == s2_use) {
1330 ct++;
1331 if (are_adjacent_refs(s1_use, s2_use)) {
1332 save_use += adjacent_profit(s1_use, s2_use);
1333 }
1334 }
1335 }
1336 }
1337 }
1338 }
1339
1340 if (ct < s1->outcnt()) save_use += unpack_cost(1);
1341 if (ct < s2->outcnt()) save_use += unpack_cost(1);
1342
1343 return MAX2(save_in, save_use);
1344 }
1345
1346 //------------------------------costs---------------------------
1347 int SuperWord::adjacent_profit(Node* s1, Node* s2) { return 2; }
1348 int SuperWord::pack_cost(int ct) { return ct; }
1349 int SuperWord::unpack_cost(int ct) { return ct; }
1350
1351 //------------------------------combine_packs---------------------------
1352 // Combine packs A and B with A.last == B.first into A.first..,A.last,B.second,..B.last
1353 void SuperWord::combine_packs() {
1354 bool changed = true;
1355 // Combine packs regardless max vector size.
1356 while (changed) {
1357 changed = false;
1358 for (int i = 0; i < _packset.length(); i++) {
1359 Node_List* p1 = _packset.at(i);
1360 if (p1 == NULL) continue;
1361 // Because of sorting we can start at i + 1
1362 for (int j = i + 1; j < _packset.length(); j++) {
1363 Node_List* p2 = _packset.at(j);
1364 if (p2 == NULL) continue;
1365 if (i == j) continue;
1366 if (p1->at(p1->size()-1) == p2->at(0)) {
1367 for (uint k = 1; k < p2->size(); k++) {
1368 p1->push(p2->at(k));
1369 }
1370 _packset.at_put(j, NULL);
1371 changed = true;
1372 }
1373 }
1374 }
1375 }
1376
1377 // Split packs which have size greater then max vector size.
1378 for (int i = 0; i < _packset.length(); i++) {
1379 Node_List* p1 = _packset.at(i);
1380 if (p1 != NULL) {
1381 BasicType bt = velt_basic_type(p1->at(0));
1382 uint max_vlen = Matcher::max_vector_size(bt); // Max elements in vector
1383 assert(is_power_of_2(max_vlen), "sanity");
1384 uint psize = p1->size();
1385 if (!is_power_of_2(psize)) {
1386 // Skip pack which can't be vector.
1387 // case1: for(...) { a[i] = i; } elements values are different (i+x)
1388 // case2: for(...) { a[i] = b[i+1]; } can't align both, load and store
1389 _packset.at_put(i, NULL);
1390 continue;
1391 }
1392 if (psize > max_vlen) {
1393 Node_List* pack = new Node_List();
1394 for (uint j = 0; j < psize; j++) {
1395 pack->push(p1->at(j));
1396 if (pack->size() >= max_vlen) {
1397 assert(is_power_of_2(pack->size()), "sanity");
1398 _packset.append(pack);
1399 pack = new Node_List();
1400 }
1401 }
1402 _packset.at_put(i, NULL);
1403 }
1404 }
1405 }
1406
1407 // Compress list.
1408 for (int i = _packset.length() - 1; i >= 0; i--) {
1409 Node_List* p1 = _packset.at(i);
1410 if (p1 == NULL) {
1411 _packset.remove_at(i);
1412 }
1413 }
1414
1415 #ifndef PRODUCT
1416 if (TraceSuperWord) {
1417 tty->print_cr("\nAfter combine_packs");
1418 print_packset();
1419 }
1420 #endif
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 NOT_PRODUCT(if (TraceSuperWordLoopUnrollAnalysis) tty->print_cr("vector loop(unroll=%d, len=%d)\n", max_vlen, max_vlen_in_bytes*BitsPerByte));
2248 // For atomic unrolled loops which are vector mapped, instigate more unrolling.
2249 cl->set_notpassed_slp();
2250 // if vector resources are limited, do not allow additional unrolling
2251 if (FLOATPRESSURE > 8) {
2252 C->set_major_progress();
2253 }
2254 cl->mark_do_unroll_only();
2255 }
2256 }
2257 }
2258
2259 if (do_reserve_copy()) {
2260 make_reversable.use_new();
2261 }
2262 NOT_PRODUCT(if(is_trace_loop_reverse()) {tty->print_cr("\n Final loop after SuperWord"); print_loop(true);})
2263 return;
2264 }
2265
2266 //------------------------------vector_opd---------------------------
2267 // Create a vector operand for the nodes in pack p for operand: in(opd_idx)
2268 Node* SuperWord::vector_opd(Node_List* p, int opd_idx) {
2269 Node* p0 = p->at(0);
2270 uint vlen = p->size();
2271 Node* opd = p0->in(opd_idx);
2272
2273 if (same_inputs(p, opd_idx)) {
2274 if (opd->is_Vector() || opd->is_LoadVector()) {
2275 assert(((opd_idx != 2) || !VectorNode::is_shift(p0)), "shift's count can't be vector");
2276 if (opd_idx == 2 && VectorNode::is_shift(p0)) {
2277 NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("shift's count can't be vector");})
2278 return NULL;
2279 }
2280 return opd; // input is matching vector
2281 }
2282 if ((opd_idx == 2) && VectorNode::is_shift(p0)) {
2283 Compile* C = _phase->C;
2284 Node* cnt = opd;
2285 // Vector instructions do not mask shift count, do it here.
2286 juint mask = (p0->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
2287 const TypeInt* t = opd->find_int_type();
2288 if (t != NULL && t->is_con()) {
2289 juint shift = t->get_con();
2290 if (shift > mask) { // Unsigned cmp
2291 cnt = ConNode::make(TypeInt::make(shift & mask));
2292 }
2293 } else {
2294 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
2295 cnt = ConNode::make(TypeInt::make(mask));
2296 _igvn.register_new_node_with_optimizer(cnt);
2297 cnt = new AndINode(opd, cnt);
2298 _igvn.register_new_node_with_optimizer(cnt);
2299 _phase->set_ctrl(cnt, _phase->get_ctrl(opd));
2300 }
2301 assert(opd->bottom_type()->isa_int(), "int type only");
2302 if (!opd->bottom_type()->isa_int()) {
2303 NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("Should be int type only");})
2304 return NULL;
2305 }
2306 // Move non constant shift count into vector register.
2307 cnt = VectorNode::shift_count(p0, cnt, vlen, velt_basic_type(p0));
2308 }
2309 if (cnt != opd) {
2310 _igvn.register_new_node_with_optimizer(cnt);
2311 _phase->set_ctrl(cnt, _phase->get_ctrl(opd));
2312 }
2313 return cnt;
2314 }
2315 assert(!opd->is_StoreVector(), "such vector is not expected here");
2316 if (opd->is_StoreVector()) {
2317 NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("StoreVector is not expected here");})
2318 return NULL;
2319 }
2320 // Convert scalar input to vector with the same number of elements as
2321 // p0's vector. Use p0's type because size of operand's container in
2322 // vector should match p0's size regardless operand's size.
2323 const Type* p0_t = velt_type(p0);
2324 VectorNode* vn = VectorNode::scalar2vector(opd, vlen, p0_t);
2325
2326 _igvn.register_new_node_with_optimizer(vn);
2327 _phase->set_ctrl(vn, _phase->get_ctrl(opd));
2328 #ifdef ASSERT
2329 if (TraceNewVectors) {
2330 tty->print("new Vector node: ");
2331 vn->dump();
2332 }
2333 #endif
2334 return vn;
2335 }
2336
2337 // Insert pack operation
2338 BasicType bt = velt_basic_type(p0);
2339 PackNode* pk = PackNode::make(opd, vlen, bt);
2340 DEBUG_ONLY( const BasicType opd_bt = opd->bottom_type()->basic_type(); )
2341
2342 for (uint i = 1; i < vlen; i++) {
2343 Node* pi = p->at(i);
2344 Node* in = pi->in(opd_idx);
2345 assert(my_pack(in) == NULL, "Should already have been unpacked");
2346 if (my_pack(in) != NULL) {
2347 NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("Should already have been unpacked");})
2348 return NULL;
2349 }
2350 assert(opd_bt == in->bottom_type()->basic_type(), "all same type");
2351 pk->add_opd(in);
2352 }
2353 _igvn.register_new_node_with_optimizer(pk);
2354 _phase->set_ctrl(pk, _phase->get_ctrl(opd));
2355 #ifdef ASSERT
2356 if (TraceNewVectors) {
2357 tty->print("new Vector node: ");
2358 pk->dump();
2359 }
2360 #endif
2361 return pk;
2362 }
2363
2364 //------------------------------insert_extracts---------------------------
2365 // If a use of pack p is not a vector use, then replace the
2366 // use with an extract operation.
2367 void SuperWord::insert_extracts(Node_List* p) {
2368 if (p->at(0)->is_Store()) return;
2369 assert(_n_idx_list.is_empty(), "empty (node,index) list");
2370
2371 // Inspect each use of each pack member. For each use that is
2372 // not a vector use, replace the use with an extract operation.
2373
2374 for (uint i = 0; i < p->size(); i++) {
2375 Node* def = p->at(i);
2376 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
2377 Node* use = def->fast_out(j);
2378 for (uint k = 0; k < use->req(); k++) {
2379 Node* n = use->in(k);
2380 if (def == n) {
2381 Node_List* u_pk = my_pack(use);
2382 if ((u_pk == NULL || !is_cmov_pack(u_pk) || use->is_CMove()) && !is_vector_use(use, k)) {
2383 _n_idx_list.push(use, k);
2384 }
2385 }
2386 }
2387 }
2388 }
2389
2390 while (_n_idx_list.is_nonempty()) {
2391 Node* use = _n_idx_list.node();
2392 int idx = _n_idx_list.index();
2393 _n_idx_list.pop();
2394 Node* def = use->in(idx);
2395
2396 if (def->is_reduction()) continue;
2397
2398 // Insert extract operation
2399 _igvn.hash_delete(def);
2400 int def_pos = alignment(def) / data_size(def);
2401
2402 Node* ex = ExtractNode::make(def, def_pos, velt_basic_type(def));
2403 _igvn.register_new_node_with_optimizer(ex);
2404 _phase->set_ctrl(ex, _phase->get_ctrl(def));
2405 _igvn.replace_input_of(use, idx, ex);
2406 _igvn._worklist.push(def);
2407
2408 bb_insert_after(ex, bb_idx(def));
2409 set_velt_type(ex, velt_type(def));
2410 }
2411 }
2412
2413 //------------------------------is_vector_use---------------------------
2414 // Is use->in(u_idx) a vector use?
2415 bool SuperWord::is_vector_use(Node* use, int u_idx) {
2416 Node_List* u_pk = my_pack(use);
2417 if (u_pk == NULL) return false;
2418 if (use->is_reduction()) return true;
2419 Node* def = use->in(u_idx);
2420 Node_List* d_pk = my_pack(def);
2421 if (d_pk == NULL) {
2422 // check for scalar promotion
2423 Node* n = u_pk->at(0)->in(u_idx);
2424 for (uint i = 1; i < u_pk->size(); i++) {
2425 if (u_pk->at(i)->in(u_idx) != n) return false;
2426 }
2427 return true;
2428 }
2429 if (u_pk->size() != d_pk->size())
2430 return false;
2431 for (uint i = 0; i < u_pk->size(); i++) {
2432 Node* ui = u_pk->at(i);
2433 Node* di = d_pk->at(i);
2434 if (ui->in(u_idx) != di || alignment(ui) != alignment(di))
2435 return false;
2436 }
2437 return true;
2438 }
2439
2440 //------------------------------construct_bb---------------------------
2441 // Construct reverse postorder list of block members
2442 bool SuperWord::construct_bb() {
2443 Node* entry = bb();
2444
2445 assert(_stk.length() == 0, "stk is empty");
2446 assert(_block.length() == 0, "block is empty");
2447 assert(_data_entry.length() == 0, "data_entry is empty");
2448 assert(_mem_slice_head.length() == 0, "mem_slice_head is empty");
2449 assert(_mem_slice_tail.length() == 0, "mem_slice_tail is empty");
2450
2451 // Find non-control nodes with no inputs from within block,
2452 // create a temporary map from node _idx to bb_idx for use
2453 // by the visited and post_visited sets,
2454 // and count number of nodes in block.
2455 int bb_ct = 0;
2456 for (uint i = 0; i < lpt()->_body.size(); i++) {
2457 Node *n = lpt()->_body.at(i);
2458 set_bb_idx(n, i); // Create a temporary map
2459 if (in_bb(n)) {
2460 if (n->is_LoadStore() || n->is_MergeMem() ||
2461 (n->is_Proj() && !n->as_Proj()->is_CFG())) {
2462 // Bailout if the loop has LoadStore, MergeMem or data Proj
2463 // nodes. Superword optimization does not work with them.
2464 return false;
2465 }
2466 bb_ct++;
2467 if (!n->is_CFG()) {
2468 bool found = false;
2469 for (uint j = 0; j < n->req(); j++) {
2470 Node* def = n->in(j);
2471 if (def && in_bb(def)) {
2472 found = true;
2473 break;
2474 }
2475 }
2476 if (!found) {
2477 assert(n != entry, "can't be entry");
2478 _data_entry.push(n);
2479 }
2480 }
2481 }
2482 }
2483
2484 // Find memory slices (head and tail)
2485 for (DUIterator_Fast imax, i = lp()->fast_outs(imax); i < imax; i++) {
2486 Node *n = lp()->fast_out(i);
2487 if (in_bb(n) && (n->is_Phi() && n->bottom_type() == Type::MEMORY)) {
2488 Node* n_tail = n->in(LoopNode::LoopBackControl);
2489 if (n_tail != n->in(LoopNode::EntryControl)) {
2490 if (!n_tail->is_Mem()) {
2491 assert(n_tail->is_Mem(), "unexpected node for memory slice: %s", n_tail->Name());
2492 return false; // Bailout
2493 }
2494 _mem_slice_head.push(n);
2495 _mem_slice_tail.push(n_tail);
2496 }
2497 }
2498 }
2499
2500 // Create an RPO list of nodes in block
2501
2502 visited_clear();
2503 post_visited_clear();
2504
2505 // Push all non-control nodes with no inputs from within block, then control entry
2506 for (int j = 0; j < _data_entry.length(); j++) {
2507 Node* n = _data_entry.at(j);
2508 visited_set(n);
2509 _stk.push(n);
2510 }
2511 visited_set(entry);
2512 _stk.push(entry);
2513
2514 // Do a depth first walk over out edges
2515 int rpo_idx = bb_ct - 1;
2516 int size;
2517 int reduction_uses = 0;
2518 while ((size = _stk.length()) > 0) {
2519 Node* n = _stk.top(); // Leave node on stack
2520 if (!visited_test_set(n)) {
2521 // forward arc in graph
2522 } else if (!post_visited_test(n)) {
2523 // cross or back arc
2524 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2525 Node *use = n->fast_out(i);
2526 if (in_bb(use) && !visited_test(use) &&
2527 // Don't go around backedge
2528 (!use->is_Phi() || n == entry)) {
2529 if (use->is_reduction()) {
2530 // First see if we can map the reduction on the given system we are on, then
2531 // make a data entry operation for each reduction we see.
2532 BasicType bt = use->bottom_type()->basic_type();
2533 if (ReductionNode::implemented(use->Opcode(), Matcher::min_vector_size(bt), bt)) {
2534 reduction_uses++;
2535 }
2536 }
2537 _stk.push(use);
2538 }
2539 }
2540 if (_stk.length() == size) {
2541 // There were no additional uses, post visit node now
2542 _stk.pop(); // Remove node from stack
2543 assert(rpo_idx >= 0, "");
2544 _block.at_put_grow(rpo_idx, n);
2545 rpo_idx--;
2546 post_visited_set(n);
2547 assert(rpo_idx >= 0 || _stk.is_empty(), "");
2548 }
2549 } else {
2550 _stk.pop(); // Remove post-visited node from stack
2551 }
2552 }//while
2553
2554 int ii_current = -1;
2555 unsigned int load_idx = (unsigned int)-1;
2556 _ii_order.clear();
2557 // Create real map of block indices for nodes
2558 for (int j = 0; j < _block.length(); j++) {
2559 Node* n = _block.at(j);
2560 set_bb_idx(n, j);
2561 if (_do_vector_loop && n->is_Load()) {
2562 if (ii_current == -1) {
2563 ii_current = _clone_map.gen(n->_idx);
2564 _ii_order.push(ii_current);
2565 load_idx = _clone_map.idx(n->_idx);
2566 } else if (_clone_map.idx(n->_idx) == load_idx && _clone_map.gen(n->_idx) != ii_current) {
2567 ii_current = _clone_map.gen(n->_idx);
2568 _ii_order.push(ii_current);
2569 }
2570 }
2571 }//for
2572
2573 // Ensure extra info is allocated.
2574 initialize_bb();
2575
2576 #ifndef PRODUCT
2577 if (_vector_loop_debug && _ii_order.length() > 0) {
2578 tty->print("SuperWord::construct_bb: List of generations: ");
2579 for (int jj = 0; jj < _ii_order.length(); ++jj) {
2580 tty->print(" %d:%d", jj, _ii_order.at(jj));
2581 }
2582 tty->print_cr(" ");
2583 }
2584 if (TraceSuperWord) {
2585 print_bb();
2586 tty->print_cr("\ndata entry nodes: %s", _data_entry.length() > 0 ? "" : "NONE");
2587 for (int m = 0; m < _data_entry.length(); m++) {
2588 tty->print("%3d ", m);
2589 _data_entry.at(m)->dump();
2590 }
2591 tty->print_cr("\nmemory slices: %s", _mem_slice_head.length() > 0 ? "" : "NONE");
2592 for (int m = 0; m < _mem_slice_head.length(); m++) {
2593 tty->print("%3d ", m); _mem_slice_head.at(m)->dump();
2594 tty->print(" "); _mem_slice_tail.at(m)->dump();
2595 }
2596 }
2597 #endif
2598 assert(rpo_idx == -1 && bb_ct == _block.length(), "all block members found");
2599 return (_mem_slice_head.length() > 0) || (reduction_uses > 0) || (_data_entry.length() > 0);
2600 }
2601
2602 //------------------------------initialize_bb---------------------------
2603 // Initialize per node info
2604 void SuperWord::initialize_bb() {
2605 Node* last = _block.at(_block.length() - 1);
2606 grow_node_info(bb_idx(last));
2607 }
2608
2609 //------------------------------bb_insert_after---------------------------
2610 // Insert n into block after pos
2611 void SuperWord::bb_insert_after(Node* n, int pos) {
2612 int n_pos = pos + 1;
2613 // Make room
2614 for (int i = _block.length() - 1; i >= n_pos; i--) {
2615 _block.at_put_grow(i+1, _block.at(i));
2616 }
2617 for (int j = _node_info.length() - 1; j >= n_pos; j--) {
2618 _node_info.at_put_grow(j+1, _node_info.at(j));
2619 }
2620 // Set value
2621 _block.at_put_grow(n_pos, n);
2622 _node_info.at_put_grow(n_pos, SWNodeInfo::initial);
2623 // Adjust map from node->_idx to _block index
2624 for (int i = n_pos; i < _block.length(); i++) {
2625 set_bb_idx(_block.at(i), i);
2626 }
2627 }
2628
2629 //------------------------------compute_max_depth---------------------------
2630 // Compute max depth for expressions from beginning of block
2631 // Use to prune search paths during test for independence.
2632 void SuperWord::compute_max_depth() {
2633 int ct = 0;
2634 bool again;
2635 do {
2636 again = false;
2637 for (int i = 0; i < _block.length(); i++) {
2638 Node* n = _block.at(i);
2639 if (!n->is_Phi()) {
2640 int d_orig = depth(n);
2641 int d_in = 0;
2642 for (DepPreds preds(n, _dg); !preds.done(); preds.next()) {
2643 Node* pred = preds.current();
2644 if (in_bb(pred)) {
2645 d_in = MAX2(d_in, depth(pred));
2646 }
2647 }
2648 if (d_in + 1 != d_orig) {
2649 set_depth(n, d_in + 1);
2650 again = true;
2651 }
2652 }
2653 }
2654 ct++;
2655 } while (again);
2656 #ifndef PRODUCT
2657 if (TraceSuperWord && Verbose)
2658 tty->print_cr("compute_max_depth iterated: %d times", ct);
2659 #endif
2660 }
2661
2662 //-------------------------compute_vector_element_type-----------------------
2663 // Compute necessary vector element type for expressions
2664 // This propagates backwards a narrower integer type when the
2665 // upper bits of the value are not needed.
2666 // Example: char a,b,c; a = b + c;
2667 // Normally the type of the add is integer, but for packed character
2668 // operations the type of the add needs to be char.
2669 void SuperWord::compute_vector_element_type() {
2670 #ifndef PRODUCT
2671 if (TraceSuperWord && Verbose)
2672 tty->print_cr("\ncompute_velt_type:");
2673 #endif
2674
2675 // Initial type
2676 for (int i = 0; i < _block.length(); i++) {
2677 Node* n = _block.at(i);
2678 set_velt_type(n, container_type(n));
2679 }
2680
2681 // Propagate integer narrowed type backwards through operations
2682 // that don't depend on higher order bits
2683 for (int i = _block.length() - 1; i >= 0; i--) {
2684 Node* n = _block.at(i);
2685 // Only integer types need be examined
2686 const Type* vtn = velt_type(n);
2687 if (vtn->basic_type() == T_INT) {
2688 uint start, end;
2689 VectorNode::vector_operands(n, &start, &end);
2690
2691 for (uint j = start; j < end; j++) {
2692 Node* in = n->in(j);
2693 // Don't propagate through a memory
2694 if (!in->is_Mem() && in_bb(in) && velt_type(in)->basic_type() == T_INT &&
2695 data_size(n) < data_size(in)) {
2696 bool same_type = true;
2697 for (DUIterator_Fast kmax, k = in->fast_outs(kmax); k < kmax; k++) {
2698 Node *use = in->fast_out(k);
2699 if (!in_bb(use) || !same_velt_type(use, n)) {
2700 same_type = false;
2701 break;
2702 }
2703 }
2704 if (same_type) {
2705 // For right shifts of small integer types (bool, byte, char, short)
2706 // we need precise information about sign-ness. Only Load nodes have
2707 // this information because Store nodes are the same for signed and
2708 // unsigned values. And any arithmetic operation after a load may
2709 // expand a value to signed Int so such right shifts can't be used
2710 // because vector elements do not have upper bits of Int.
2711 const Type* vt = vtn;
2712 if (VectorNode::is_shift(in)) {
2713 Node* load = in->in(1);
2714 if (load->is_Load() && in_bb(load) && (velt_type(load)->basic_type() == T_INT)) {
2715 vt = velt_type(load);
2716 } else if (in->Opcode() != Op_LShiftI) {
2717 // Widen type to Int to avoid creation of right shift vector
2718 // (align + data_size(s1) check in stmts_can_pack() will fail).
2719 // Note, left shifts work regardless type.
2720 vt = TypeInt::INT;
2721 }
2722 }
2723 set_velt_type(in, vt);
2724 }
2725 }
2726 }
2727 }
2728 }
2729 #ifndef PRODUCT
2730 if (TraceSuperWord && Verbose) {
2731 for (int i = 0; i < _block.length(); i++) {
2732 Node* n = _block.at(i);
2733 velt_type(n)->dump();
2734 tty->print("\t");
2735 n->dump();
2736 }
2737 }
2738 #endif
2739 }
2740
2741 //------------------------------memory_alignment---------------------------
2742 // Alignment within a vector memory reference
2743 int SuperWord::memory_alignment(MemNode* s, int iv_adjust) {
2744 #ifndef PRODUCT
2745 if(TraceSuperWord && Verbose) {
2746 tty->print("SuperWord::memory_alignment within a vector memory reference for %d: ", s->_idx); s->dump();
2747 }
2748 #endif
2749 NOT_PRODUCT(SWPointer::Tracer::Depth ddd(0);)
2750 SWPointer p(s, this, NULL, false);
2751 if (!p.valid()) {
2752 NOT_PRODUCT(if(is_trace_alignment()) tty->print("SWPointer::memory_alignment: SWPointer p invalid, return bottom_align");)
2753 return bottom_align;
2754 }
2755 int vw = vector_width_in_bytes(s);
2756 if (vw < 2) {
2757 NOT_PRODUCT(if(is_trace_alignment()) tty->print_cr("SWPointer::memory_alignment: vector_width_in_bytes < 2, return bottom_align");)
2758 return bottom_align; // No vectors for this type
2759 }
2760 int offset = p.offset_in_bytes();
2761 offset += iv_adjust*p.memory_size();
2762 int off_rem = offset % vw;
2763 int off_mod = off_rem >= 0 ? off_rem : off_rem + vw;
2764 NOT_PRODUCT(if(TraceSuperWord && Verbose) tty->print_cr("SWPointer::memory_alignment: off_rem = %d, off_mod = %d", off_rem, off_mod);)
2765 return off_mod;
2766 }
2767
2768 //---------------------------container_type---------------------------
2769 // Smallest type containing range of values
2770 const Type* SuperWord::container_type(Node* n) {
2771 if (n->is_Mem()) {
2772 BasicType bt = n->as_Mem()->memory_type();
2773 if (n->is_Store() && (bt == T_CHAR)) {
2774 // Use T_SHORT type instead of T_CHAR for stored values because any
2775 // preceding arithmetic operation extends values to signed Int.
2776 bt = T_SHORT;
2777 }
2778 if (n->Opcode() == Op_LoadUB) {
2779 // Adjust type for unsigned byte loads, it is important for right shifts.
2780 // T_BOOLEAN is used because there is no basic type representing type
2781 // TypeInt::UBYTE. Use of T_BOOLEAN for vectors is fine because only
2782 // size (one byte) and sign is important.
2783 bt = T_BOOLEAN;
2784 }
2785 return Type::get_const_basic_type(bt);
2786 }
2787 const Type* t = _igvn.type(n);
2788 if (t->basic_type() == T_INT) {
2789 // A narrow type of arithmetic operations will be determined by
2790 // propagating the type of memory operations.
2791 return TypeInt::INT;
2792 }
2793 return t;
2794 }
2795
2796 bool SuperWord::same_velt_type(Node* n1, Node* n2) {
2797 const Type* vt1 = velt_type(n1);
2798 const Type* vt2 = velt_type(n2);
2799 if (vt1->basic_type() == T_INT && vt2->basic_type() == T_INT) {
2800 // Compare vectors element sizes for integer types.
2801 return data_size(n1) == data_size(n2);
2802 }
2803 return vt1 == vt2;
2804 }
2805
2806 //------------------------------in_packset---------------------------
2807 // Are s1 and s2 in a pack pair and ordered as s1,s2?
2808 bool SuperWord::in_packset(Node* s1, Node* s2) {
2809 for (int i = 0; i < _packset.length(); i++) {
2810 Node_List* p = _packset.at(i);
2811 assert(p->size() == 2, "must be");
2812 if (p->at(0) == s1 && p->at(p->size()-1) == s2) {
2813 return true;
2814 }
2815 }
2816 return false;
2817 }
2818
2819 //------------------------------in_pack---------------------------
2820 // Is s in pack p?
2821 Node_List* SuperWord::in_pack(Node* s, Node_List* p) {
2822 for (uint i = 0; i < p->size(); i++) {
2823 if (p->at(i) == s) {
2824 return p;
2825 }
2826 }
2827 return NULL;
2828 }
2829
2830 //------------------------------remove_pack_at---------------------------
2831 // Remove the pack at position pos in the packset
2832 void SuperWord::remove_pack_at(int pos) {
2833 Node_List* p = _packset.at(pos);
2834 for (uint i = 0; i < p->size(); i++) {
2835 Node* s = p->at(i);
2836 set_my_pack(s, NULL);
2837 }
2838 _packset.remove_at(pos);
2839 }
2840
2841 void SuperWord::packset_sort(int n) {
2842 // simple bubble sort so that we capitalize with O(n) when its already sorted
2843 while (n != 0) {
2844 bool swapped = false;
2845 for (int i = 1; i < n; i++) {
2846 Node_List* q_low = _packset.at(i-1);
2847 Node_List* q_i = _packset.at(i);
2848
2849 // only swap when we find something to swap
2850 if (alignment(q_low->at(0)) > alignment(q_i->at(0))) {
2851 Node_List* t = q_i;
2852 *(_packset.adr_at(i)) = q_low;
2853 *(_packset.adr_at(i-1)) = q_i;
2854 swapped = true;
2855 }
2856 }
2857 if (swapped == false) break;
2858 n--;
2859 }
2860 }
2861
2862 //------------------------------executed_first---------------------------
2863 // Return the node executed first in pack p. Uses the RPO block list
2864 // to determine order.
2865 Node* SuperWord::executed_first(Node_List* p) {
2866 Node* n = p->at(0);
2867 int n_rpo = bb_idx(n);
2868 for (uint i = 1; i < p->size(); i++) {
2869 Node* s = p->at(i);
2870 int s_rpo = bb_idx(s);
2871 if (s_rpo < n_rpo) {
2872 n = s;
2873 n_rpo = s_rpo;
2874 }
2875 }
2876 return n;
2877 }
2878
2879 //------------------------------executed_last---------------------------
2880 // Return the node executed last in pack p.
2881 Node* SuperWord::executed_last(Node_List* p) {
2882 Node* n = p->at(0);
2883 int n_rpo = bb_idx(n);
2884 for (uint i = 1; i < p->size(); i++) {
2885 Node* s = p->at(i);
2886 int s_rpo = bb_idx(s);
2887 if (s_rpo > n_rpo) {
2888 n = s;
2889 n_rpo = s_rpo;
2890 }
2891 }
2892 return n;
2893 }
2894
2895 LoadNode::ControlDependency SuperWord::control_dependency(Node_List* p) {
2896 LoadNode::ControlDependency dep = LoadNode::DependsOnlyOnTest;
2897 for (uint i = 0; i < p->size(); i++) {
2898 Node* n = p->at(i);
2899 assert(n->is_Load(), "only meaningful for loads");
2900 if (!n->depends_only_on_test()) {
2901 dep = LoadNode::Pinned;
2902 }
2903 }
2904 return dep;
2905 }
2906
2907
2908 //----------------------------align_initial_loop_index---------------------------
2909 // Adjust pre-loop limit so that in main loop, a load/store reference
2910 // to align_to_ref will be a position zero in the vector.
2911 // (iv + k) mod vector_align == 0
2912 void SuperWord::align_initial_loop_index(MemNode* align_to_ref) {
2913 CountedLoopNode *main_head = lp()->as_CountedLoop();
2914 assert(main_head->is_main_loop(), "");
2915 CountedLoopEndNode* pre_end = get_pre_loop_end(main_head);
2916 assert(pre_end != NULL, "we must have a correct pre-loop");
2917 Node *pre_opaq1 = pre_end->limit();
2918 assert(pre_opaq1->Opcode() == Op_Opaque1, "");
2919 Opaque1Node *pre_opaq = (Opaque1Node*)pre_opaq1;
2920 Node *lim0 = pre_opaq->in(1);
2921
2922 // Where we put new limit calculations
2923 Node *pre_ctrl = pre_end->loopnode()->in(LoopNode::EntryControl);
2924
2925 // Ensure the original loop limit is available from the
2926 // pre-loop Opaque1 node.
2927 Node *orig_limit = pre_opaq->original_loop_limit();
2928 assert(orig_limit != NULL && _igvn.type(orig_limit) != Type::TOP, "");
2929
2930 SWPointer align_to_ref_p(align_to_ref, this, NULL, false);
2931 assert(align_to_ref_p.valid(), "sanity");
2932
2933 // Given:
2934 // lim0 == original pre loop limit
2935 // V == v_align (power of 2)
2936 // invar == extra invariant piece of the address expression
2937 // e == offset [ +/- invar ]
2938 //
2939 // When reassociating expressions involving '%' the basic rules are:
2940 // (a - b) % k == 0 => a % k == b % k
2941 // and:
2942 // (a + b) % k == 0 => a % k == (k - b) % k
2943 //
2944 // For stride > 0 && scale > 0,
2945 // Derive the new pre-loop limit "lim" such that the two constraints:
2946 // (1) lim = lim0 + N (where N is some positive integer < V)
2947 // (2) (e + lim) % V == 0
2948 // are true.
2949 //
2950 // Substituting (1) into (2),
2951 // (e + lim0 + N) % V == 0
2952 // solve for N:
2953 // N = (V - (e + lim0)) % V
2954 // substitute back into (1), so that new limit
2955 // lim = lim0 + (V - (e + lim0)) % V
2956 //
2957 // For stride > 0 && scale < 0
2958 // Constraints:
2959 // lim = lim0 + N
2960 // (e - lim) % V == 0
2961 // Solving for lim:
2962 // (e - lim0 - N) % V == 0
2963 // N = (e - lim0) % V
2964 // lim = lim0 + (e - lim0) % V
2965 //
2966 // For stride < 0 && scale > 0
2967 // Constraints:
2968 // lim = lim0 - N
2969 // (e + lim) % V == 0
2970 // Solving for lim:
2971 // (e + lim0 - N) % V == 0
2972 // N = (e + lim0) % V
2973 // lim = lim0 - (e + lim0) % V
2974 //
2975 // For stride < 0 && scale < 0
2976 // Constraints:
2977 // lim = lim0 - N
2978 // (e - lim) % V == 0
2979 // Solving for lim:
2980 // (e - lim0 + N) % V == 0
2981 // N = (V - (e - lim0)) % V
2982 // lim = lim0 - (V - (e - lim0)) % V
2983
2984 int vw = vector_width_in_bytes(align_to_ref);
2985 int stride = iv_stride();
2986 int scale = align_to_ref_p.scale_in_bytes();
2987 int elt_size = align_to_ref_p.memory_size();
2988 int v_align = vw / elt_size;
2989 assert(v_align > 1, "sanity");
2990 int offset = align_to_ref_p.offset_in_bytes() / elt_size;
2991 Node *offsn = _igvn.intcon(offset);
2992
2993 Node *e = offsn;
2994 if (align_to_ref_p.invar() != NULL) {
2995 // incorporate any extra invariant piece producing (offset +/- invar) >>> log2(elt)
2996 Node* log2_elt = _igvn.intcon(exact_log2(elt_size));
2997 Node* aref = new URShiftINode(align_to_ref_p.invar(), log2_elt);
2998 _igvn.register_new_node_with_optimizer(aref);
2999 _phase->set_ctrl(aref, pre_ctrl);
3000 if (align_to_ref_p.negate_invar()) {
3001 e = new SubINode(e, aref);
3002 } else {
3003 e = new AddINode(e, aref);
3004 }
3005 _igvn.register_new_node_with_optimizer(e);
3006 _phase->set_ctrl(e, pre_ctrl);
3007 }
3008 if (vw > ObjectAlignmentInBytes) {
3009 // incorporate base e +/- base && Mask >>> log2(elt)
3010 Node* xbase = new CastP2XNode(NULL, align_to_ref_p.base());
3011 _igvn.register_new_node_with_optimizer(xbase);
3012 #ifdef _LP64
3013 xbase = new ConvL2INode(xbase);
3014 _igvn.register_new_node_with_optimizer(xbase);
3015 #endif
3016 Node* mask = _igvn.intcon(vw-1);
3017 Node* masked_xbase = new AndINode(xbase, mask);
3018 _igvn.register_new_node_with_optimizer(masked_xbase);
3019 Node* log2_elt = _igvn.intcon(exact_log2(elt_size));
3020 Node* bref = new URShiftINode(masked_xbase, log2_elt);
3021 _igvn.register_new_node_with_optimizer(bref);
3022 _phase->set_ctrl(bref, pre_ctrl);
3023 e = new AddINode(e, bref);
3024 _igvn.register_new_node_with_optimizer(e);
3025 _phase->set_ctrl(e, pre_ctrl);
3026 }
3027
3028 // compute e +/- lim0
3029 if (scale < 0) {
3030 e = new SubINode(e, lim0);
3031 } else {
3032 e = new AddINode(e, lim0);
3033 }
3034 _igvn.register_new_node_with_optimizer(e);
3035 _phase->set_ctrl(e, pre_ctrl);
3036
3037 if (stride * scale > 0) {
3038 // compute V - (e +/- lim0)
3039 Node* va = _igvn.intcon(v_align);
3040 e = new SubINode(va, e);
3041 _igvn.register_new_node_with_optimizer(e);
3042 _phase->set_ctrl(e, pre_ctrl);
3043 }
3044 // compute N = (exp) % V
3045 Node* va_msk = _igvn.intcon(v_align - 1);
3046 Node* N = new AndINode(e, va_msk);
3047 _igvn.register_new_node_with_optimizer(N);
3048 _phase->set_ctrl(N, pre_ctrl);
3049
3050 // substitute back into (1), so that new limit
3051 // lim = lim0 + N
3052 Node* lim;
3053 if (stride < 0) {
3054 lim = new SubINode(lim0, N);
3055 } else {
3056 lim = new AddINode(lim0, N);
3057 }
3058 _igvn.register_new_node_with_optimizer(lim);
3059 _phase->set_ctrl(lim, pre_ctrl);
3060 Node* constrained =
3061 (stride > 0) ? (Node*) new MinINode(lim, orig_limit)
3062 : (Node*) new MaxINode(lim, orig_limit);
3063 _igvn.register_new_node_with_optimizer(constrained);
3064 _phase->set_ctrl(constrained, pre_ctrl);
3065 _igvn.hash_delete(pre_opaq);
3066 pre_opaq->set_req(1, constrained);
3067 }
3068
3069 //----------------------------get_pre_loop_end---------------------------
3070 // Find pre loop end from main loop. Returns null if none.
3071 CountedLoopEndNode* SuperWord::get_pre_loop_end(CountedLoopNode* cl) {
3072 Node* ctrl = cl->in(LoopNode::EntryControl);
3073 if (!ctrl->is_IfTrue() && !ctrl->is_IfFalse()) return NULL;
3074 Node* iffm = ctrl->in(0);
3075 if (!iffm->is_If()) return NULL;
3076 Node* bolzm = iffm->in(1);
3077 if (!bolzm->is_Bool()) return NULL;
3078 Node* cmpzm = bolzm->in(1);
3079 if (!cmpzm->is_Cmp()) return NULL;
3080 Node* opqzm = cmpzm->in(2);
3081 // Can not optimize a loop if zero-trip Opaque1 node is optimized
3082 // away and then another round of loop opts attempted.
3083 if (opqzm->Opcode() != Op_Opaque1) {
3084 return NULL;
3085 }
3086 Node* p_f = iffm->in(0);
3087 if (!p_f->is_IfFalse()) return NULL;
3088 if (!p_f->in(0)->is_CountedLoopEnd()) return NULL;
3089 CountedLoopEndNode* pre_end = p_f->in(0)->as_CountedLoopEnd();
3090 CountedLoopNode* loop_node = pre_end->loopnode();
3091 if (loop_node == NULL || !loop_node->is_pre_loop()) return NULL;
3092 return pre_end;
3093 }
3094
3095
3096 //------------------------------init---------------------------
3097 void SuperWord::init() {
3098 _dg.init();
3099 _packset.clear();
3100 _disjoint_ptrs.clear();
3101 _block.clear();
3102 _data_entry.clear();
3103 _mem_slice_head.clear();
3104 _mem_slice_tail.clear();
3105 _iteration_first.clear();
3106 _iteration_last.clear();
3107 _node_info.clear();
3108 _align_to_ref = NULL;
3109 _lpt = NULL;
3110 _lp = NULL;
3111 _bb = NULL;
3112 _iv = NULL;
3113 _race_possible = 0;
3114 _early_return = false;
3115 _num_work_vecs = 0;
3116 _num_reductions = 0;
3117 }
3118
3119 //------------------------------restart---------------------------
3120 void SuperWord::restart() {
3121 _dg.init();
3122 _packset.clear();
3123 _disjoint_ptrs.clear();
3124 _block.clear();
3125 _data_entry.clear();
3126 _mem_slice_head.clear();
3127 _mem_slice_tail.clear();
3128 _node_info.clear();
3129 }
3130
3131 //------------------------------print_packset---------------------------
3132 void SuperWord::print_packset() {
3133 #ifndef PRODUCT
3134 tty->print_cr("packset");
3135 for (int i = 0; i < _packset.length(); i++) {
3136 tty->print_cr("Pack: %d", i);
3137 Node_List* p = _packset.at(i);
3138 print_pack(p);
3139 }
3140 #endif
3141 }
3142
3143 //------------------------------print_pack---------------------------
3144 void SuperWord::print_pack(Node_List* p) {
3145 for (uint i = 0; i < p->size(); i++) {
3146 print_stmt(p->at(i));
3147 }
3148 }
3149
3150 //------------------------------print_bb---------------------------
3151 void SuperWord::print_bb() {
3152 #ifndef PRODUCT
3153 tty->print_cr("\nBlock");
3154 for (int i = 0; i < _block.length(); i++) {
3155 Node* n = _block.at(i);
3156 tty->print("%d ", i);
3157 if (n) {
3158 n->dump();
3159 }
3160 }
3161 #endif
3162 }
3163
3164 //------------------------------print_stmt---------------------------
3165 void SuperWord::print_stmt(Node* s) {
3166 #ifndef PRODUCT
3167 tty->print(" align: %d \t", alignment(s));
3168 s->dump();
3169 #endif
3170 }
3171
3172 //------------------------------blank---------------------------
3173 char* SuperWord::blank(uint depth) {
3174 static char blanks[101];
3175 assert(depth < 101, "too deep");
3176 for (uint i = 0; i < depth; i++) blanks[i] = ' ';
3177 blanks[depth] = '\0';
3178 return blanks;
3179 }
3180
3181
3182 //==============================SWPointer===========================
3183 #ifndef PRODUCT
3184 int SWPointer::Tracer::_depth = 0;
3185 #endif
3186 //----------------------------SWPointer------------------------
3187 SWPointer::SWPointer(MemNode* mem, SuperWord* slp, Node_Stack *nstack, bool analyze_only) :
3188 _mem(mem), _slp(slp), _base(NULL), _adr(NULL),
3189 _scale(0), _offset(0), _invar(NULL), _negate_invar(false),
3190 _nstack(nstack), _analyze_only(analyze_only),
3191 _stack_idx(0)
3192 #ifndef PRODUCT
3193 , _tracer(slp)
3194 #endif
3195 {
3196 NOT_PRODUCT(_tracer.ctor_1(mem);)
3197
3198 Node* adr = mem->in(MemNode::Address);
3199 if (!adr->is_AddP()) {
3200 assert(!valid(), "too complex");
3201 return;
3202 }
3203 // Match AddP(base, AddP(ptr, k*iv [+ invariant]), constant)
3204 Node* base = adr->in(AddPNode::Base);
3205 // The base address should be loop invariant
3206 if (!invariant(base)) {
3207 assert(!valid(), "base address is loop variant");
3208 return;
3209 }
3210 //unsafe reference could not be aligned appropriately without runtime checking
3211 if (base == NULL || base->bottom_type() == Type::TOP) {
3212 assert(!valid(), "unsafe access");
3213 return;
3214 }
3215
3216 NOT_PRODUCT(if(_slp->is_trace_alignment()) _tracer.store_depth();)
3217 NOT_PRODUCT(_tracer.ctor_2(adr);)
3218
3219 int i;
3220 for (i = 0; i < 3; i++) {
3221 NOT_PRODUCT(_tracer.ctor_3(adr, i);)
3222
3223 if (!scaled_iv_plus_offset(adr->in(AddPNode::Offset))) {
3224 assert(!valid(), "too complex");
3225 return;
3226 }
3227 adr = adr->in(AddPNode::Address);
3228 NOT_PRODUCT(_tracer.ctor_4(adr, i);)
3229
3230 if (base == adr || !adr->is_AddP()) {
3231 NOT_PRODUCT(_tracer.ctor_5(adr, base, i);)
3232 break; // stop looking at addp's
3233 }
3234 }
3235 NOT_PRODUCT(if(_slp->is_trace_alignment()) _tracer.restore_depth();)
3236 NOT_PRODUCT(_tracer.ctor_6(mem);)
3237
3238 _base = base;
3239 _adr = adr;
3240 assert(valid(), "Usable");
3241 }
3242
3243 // Following is used to create a temporary object during
3244 // the pattern match of an address expression.
3245 SWPointer::SWPointer(SWPointer* p) :
3246 _mem(p->_mem), _slp(p->_slp), _base(NULL), _adr(NULL),
3247 _scale(0), _offset(0), _invar(NULL), _negate_invar(false),
3248 _nstack(p->_nstack), _analyze_only(p->_analyze_only),
3249 _stack_idx(p->_stack_idx)
3250 #ifndef PRODUCT
3251 , _tracer(p->_slp)
3252 #endif
3253 {}
3254
3255
3256 bool SWPointer::invariant(Node* n) {
3257 NOT_PRODUCT(Tracer::Depth dd;)
3258 Node *n_c = phase()->get_ctrl(n);
3259 NOT_PRODUCT(_tracer.invariant_1(n, n_c);)
3260 return !lpt()->is_member(phase()->get_loop(n_c));
3261 }
3262 //------------------------scaled_iv_plus_offset--------------------
3263 // Match: k*iv + offset
3264 // where: k is a constant that maybe zero, and
3265 // offset is (k2 [+/- invariant]) where k2 maybe zero and invariant is optional
3266 bool SWPointer::scaled_iv_plus_offset(Node* n) {
3267 NOT_PRODUCT(Tracer::Depth ddd;)
3268 NOT_PRODUCT(_tracer.scaled_iv_plus_offset_1(n);)
3269
3270 if (scaled_iv(n)) {
3271 NOT_PRODUCT(_tracer.scaled_iv_plus_offset_2(n);)
3272 return true;
3273 }
3274
3275 if (offset_plus_k(n)) {
3276 NOT_PRODUCT(_tracer.scaled_iv_plus_offset_3(n);)
3277 return true;
3278 }
3279
3280 int opc = n->Opcode();
3281 if (opc == Op_AddI) {
3282 if (scaled_iv(n->in(1)) && offset_plus_k(n->in(2))) {
3283 NOT_PRODUCT(_tracer.scaled_iv_plus_offset_4(n);)
3284 return true;
3285 }
3286 if (scaled_iv(n->in(2)) && offset_plus_k(n->in(1))) {
3287 NOT_PRODUCT(_tracer.scaled_iv_plus_offset_5(n);)
3288 return true;
3289 }
3290 } else if (opc == Op_SubI) {
3291 if (scaled_iv(n->in(1)) && offset_plus_k(n->in(2), true)) {
3292 NOT_PRODUCT(_tracer.scaled_iv_plus_offset_6(n);)
3293 return true;
3294 }
3295 if (scaled_iv(n->in(2)) && offset_plus_k(n->in(1))) {
3296 _scale *= -1;
3297 NOT_PRODUCT(_tracer.scaled_iv_plus_offset_7(n);)
3298 return true;
3299 }
3300 }
3301
3302 NOT_PRODUCT(_tracer.scaled_iv_plus_offset_8(n);)
3303 return false;
3304 }
3305
3306 //----------------------------scaled_iv------------------------
3307 // Match: k*iv where k is a constant that's not zero
3308 bool SWPointer::scaled_iv(Node* n) {
3309 NOT_PRODUCT(Tracer::Depth ddd;)
3310 NOT_PRODUCT(_tracer.scaled_iv_1(n);)
3311
3312 if (_scale != 0) { // already found a scale
3313 NOT_PRODUCT(_tracer.scaled_iv_2(n, _scale);)
3314 return false;
3315 }
3316
3317 if (n == iv()) {
3318 _scale = 1;
3319 NOT_PRODUCT(_tracer.scaled_iv_3(n, _scale);)
3320 return true;
3321 }
3322 if (_analyze_only && (invariant(n) == false)) {
3323 _nstack->push(n, _stack_idx++);
3324 }
3325
3326 int opc = n->Opcode();
3327 if (opc == Op_MulI) {
3328 if (n->in(1) == iv() && n->in(2)->is_Con()) {
3329 _scale = n->in(2)->get_int();
3330 NOT_PRODUCT(_tracer.scaled_iv_4(n, _scale);)
3331 return true;
3332 } else if (n->in(2) == iv() && n->in(1)->is_Con()) {
3333 _scale = n->in(1)->get_int();
3334 NOT_PRODUCT(_tracer.scaled_iv_5(n, _scale);)
3335 return true;
3336 }
3337 } else if (opc == Op_LShiftI) {
3338 if (n->in(1) == iv() && n->in(2)->is_Con()) {
3339 _scale = 1 << n->in(2)->get_int();
3340 NOT_PRODUCT(_tracer.scaled_iv_6(n, _scale);)
3341 return true;
3342 }
3343 } else if (opc == Op_ConvI2L) {
3344 if (scaled_iv_plus_offset(n->in(1))) {
3345 NOT_PRODUCT(_tracer.scaled_iv_7(n);)
3346 return true;
3347 }
3348 } else if (opc == Op_LShiftL) {
3349 if (!has_iv() && _invar == NULL) {
3350 // Need to preserve the current _offset value, so
3351 // create a temporary object for this expression subtree.
3352 // Hacky, so should re-engineer the address pattern match.
3353 NOT_PRODUCT(Tracer::Depth dddd;)
3354 SWPointer tmp(this);
3355 NOT_PRODUCT(_tracer.scaled_iv_8(n, &tmp);)
3356
3357 if (tmp.scaled_iv_plus_offset(n->in(1))) {
3358 if (tmp._invar == NULL || _slp->do_vector_loop()) {
3359 int mult = 1 << n->in(2)->get_int();
3360 _scale = tmp._scale * mult;
3361 _offset += tmp._offset * mult;
3362 NOT_PRODUCT(_tracer.scaled_iv_9(n, _scale, _offset, mult);)
3363 return true;
3364 }
3365 }
3366 }
3367 }
3368 NOT_PRODUCT(_tracer.scaled_iv_10(n);)
3369 return false;
3370 }
3371
3372 //----------------------------offset_plus_k------------------------
3373 // Match: offset is (k [+/- invariant])
3374 // where k maybe zero and invariant is optional, but not both.
3375 bool SWPointer::offset_plus_k(Node* n, bool negate) {
3376 NOT_PRODUCT(Tracer::Depth ddd;)
3377 NOT_PRODUCT(_tracer.offset_plus_k_1(n);)
3378
3379 int opc = n->Opcode();
3380 if (opc == Op_ConI) {
3381 _offset += negate ? -(n->get_int()) : n->get_int();
3382 NOT_PRODUCT(_tracer.offset_plus_k_2(n, _offset);)
3383 return true;
3384 } else if (opc == Op_ConL) {
3385 // Okay if value fits into an int
3386 const TypeLong* t = n->find_long_type();
3387 if (t->higher_equal(TypeLong::INT)) {
3388 jlong loff = n->get_long();
3389 jint off = (jint)loff;
3390 _offset += negate ? -off : loff;
3391 NOT_PRODUCT(_tracer.offset_plus_k_3(n, _offset);)
3392 return true;
3393 }
3394 NOT_PRODUCT(_tracer.offset_plus_k_4(n);)
3395 return false;
3396 }
3397 if (_invar != NULL) { // already has an invariant
3398 NOT_PRODUCT(_tracer.offset_plus_k_5(n, _invar);)
3399 return false;
3400 }
3401
3402 if (_analyze_only && (invariant(n) == false)) {
3403 _nstack->push(n, _stack_idx++);
3404 }
3405 if (opc == Op_AddI) {
3406 if (n->in(2)->is_Con() && invariant(n->in(1))) {
3407 _negate_invar = negate;
3408 _invar = n->in(1);
3409 _offset += negate ? -(n->in(2)->get_int()) : n->in(2)->get_int();
3410 NOT_PRODUCT(_tracer.offset_plus_k_6(n, _invar, _negate_invar, _offset);)
3411 return true;
3412 } else if (n->in(1)->is_Con() && invariant(n->in(2))) {
3413 _offset += negate ? -(n->in(1)->get_int()) : n->in(1)->get_int();
3414 _negate_invar = negate;
3415 _invar = n->in(2);
3416 NOT_PRODUCT(_tracer.offset_plus_k_7(n, _invar, _negate_invar, _offset);)
3417 return true;
3418 }
3419 }
3420 if (opc == Op_SubI) {
3421 if (n->in(2)->is_Con() && invariant(n->in(1))) {
3422 _negate_invar = negate;
3423 _invar = n->in(1);
3424 _offset += !negate ? -(n->in(2)->get_int()) : n->in(2)->get_int();
3425 NOT_PRODUCT(_tracer.offset_plus_k_8(n, _invar, _negate_invar, _offset);)
3426 return true;
3427 } else if (n->in(1)->is_Con() && invariant(n->in(2))) {
3428 _offset += negate ? -(n->in(1)->get_int()) : n->in(1)->get_int();
3429 _negate_invar = !negate;
3430 _invar = n->in(2);
3431 NOT_PRODUCT(_tracer.offset_plus_k_9(n, _invar, _negate_invar, _offset);)
3432 return true;
3433 }
3434 }
3435 if (invariant(n)) {
3436 if (opc == Op_ConvI2L) {
3437 n = n->in(1);
3438 }
3439 _negate_invar = negate;
3440 _invar = n;
3441 NOT_PRODUCT(_tracer.offset_plus_k_10(n, _invar, _negate_invar, _offset);)
3442 return true;
3443 }
3444
3445 NOT_PRODUCT(_tracer.offset_plus_k_11(n);)
3446 return false;
3447 }
3448
3449 //----------------------------print------------------------
3450 void SWPointer::print() {
3451 #ifndef PRODUCT
3452 tty->print("base: %d adr: %d scale: %d offset: %d invar: %c%d\n",
3453 _base != NULL ? _base->_idx : 0,
3454 _adr != NULL ? _adr->_idx : 0,
3455 _scale, _offset,
3456 _negate_invar?'-':'+',
3457 _invar != NULL ? _invar->_idx : 0);
3458 #endif
3459 }
3460
3461 //----------------------------tracing------------------------
3462 #ifndef PRODUCT
3463 void SWPointer::Tracer::print_depth() {
3464 for (int ii = 0; ii<_depth; ++ii) tty->print(" ");
3465 }
3466
3467 void SWPointer::Tracer::ctor_1 (Node* mem) {
3468 if(_slp->is_trace_alignment()) {
3469 print_depth(); tty->print(" %d SWPointer::SWPointer: start alignment analysis", mem->_idx); mem->dump();
3470 }
3471 }
3472
3473 void SWPointer::Tracer::ctor_2(Node* adr) {
3474 if(_slp->is_trace_alignment()) {
3475 //store_depth();
3476 inc_depth();
3477 print_depth(); tty->print(" %d (adr) SWPointer::SWPointer: ", adr->_idx); adr->dump();
3478 inc_depth();
3479 print_depth(); tty->print(" %d (base) SWPointer::SWPointer: ", adr->in(AddPNode::Base)->_idx); adr->in(AddPNode::Base)->dump();
3480 }
3481 }
3482
3483 void SWPointer::Tracer::ctor_3(Node* adr, int i) {
3484 if(_slp->is_trace_alignment()) {
3485 inc_depth();
3486 Node* offset = adr->in(AddPNode::Offset);
3487 print_depth(); tty->print(" %d (offset) SWPointer::SWPointer: i = %d: ", offset->_idx, i); offset->dump();
3488 }
3489 }
3490
3491 void SWPointer::Tracer::ctor_4(Node* adr, int i) {
3492 if(_slp->is_trace_alignment()) {
3493 inc_depth();
3494 print_depth(); tty->print(" %d (adr) SWPointer::SWPointer: i = %d: ", adr->_idx, i); adr->dump();
3495 }
3496 }
3497
3498 void SWPointer::Tracer::ctor_5(Node* adr, Node* base, int i) {
3499 if(_slp->is_trace_alignment()) {
3500 inc_depth();
3501 if (base == adr) {
3502 print_depth(); tty->print_cr(" \\ %d (adr) == %d (base) SWPointer::SWPointer: breaking analysis at i = %d", adr->_idx, base->_idx, i);
3503 } else if (!adr->is_AddP()) {
3504 print_depth(); tty->print_cr(" \\ %d (adr) is NOT Addp SWPointer::SWPointer: breaking analysis at i = %d", adr->_idx, i);
3505 }
3506 }
3507 }
3508
3509 void SWPointer::Tracer::ctor_6(Node* mem) {
3510 if(_slp->is_trace_alignment()) {
3511 //restore_depth();
3512 print_depth(); tty->print_cr(" %d (adr) SWPointer::SWPointer: stop analysis", mem->_idx);
3513 }
3514 }
3515
3516 void SWPointer::Tracer::invariant_1(Node *n, Node *n_c) {
3517 if (_slp->do_vector_loop() && _slp->is_debug() && _slp->_lpt->is_member(_slp->_phase->get_loop(n_c)) != (int)_slp->in_bb(n)) {
3518 int is_member = _slp->_lpt->is_member(_slp->_phase->get_loop(n_c));
3519 int in_bb = _slp->in_bb(n);
3520 print_depth(); tty->print(" \\ "); tty->print_cr(" %d SWPointer::invariant conditions differ: n_c %d", n->_idx, n_c->_idx);
3521 print_depth(); tty->print(" \\ "); tty->print_cr("is_member %d, in_bb %d", is_member, in_bb);
3522 print_depth(); tty->print(" \\ "); n->dump();
3523 print_depth(); tty->print(" \\ "); n_c->dump();
3524 }
3525 }
3526
3527 void SWPointer::Tracer::scaled_iv_plus_offset_1(Node* n) {
3528 if(_slp->is_trace_alignment()) {
3529 print_depth(); tty->print(" %d SWPointer::scaled_iv_plus_offset testing node: ", n->_idx);
3530 n->dump();
3531 }
3532 }
3533
3534 void SWPointer::Tracer::scaled_iv_plus_offset_2(Node* n) {
3535 if(_slp->is_trace_alignment()) {
3536 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: PASSED", n->_idx);
3537 }
3538 }
3539
3540 void SWPointer::Tracer::scaled_iv_plus_offset_3(Node* n) {
3541 if(_slp->is_trace_alignment()) {
3542 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: PASSED", n->_idx);
3543 }
3544 }
3545
3546 void SWPointer::Tracer::scaled_iv_plus_offset_4(Node* n) {
3547 if(_slp->is_trace_alignment()) {
3548 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: Op_AddI PASSED", n->_idx);
3549 print_depth(); tty->print(" \\ %d SWPointer::scaled_iv_plus_offset: in(1) is scaled_iv: ", n->in(1)->_idx); n->in(1)->dump();
3550 print_depth(); tty->print(" \\ %d SWPointer::scaled_iv_plus_offset: in(2) is offset_plus_k: ", n->in(2)->_idx); n->in(2)->dump();
3551 }
3552 }
3553
3554 void SWPointer::Tracer::scaled_iv_plus_offset_5(Node* n) {
3555 if(_slp->is_trace_alignment()) {
3556 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: Op_AddI PASSED", n->_idx);
3557 print_depth(); tty->print(" \\ %d SWPointer::scaled_iv_plus_offset: in(2) is scaled_iv: ", n->in(2)->_idx); n->in(2)->dump();
3558 print_depth(); tty->print(" \\ %d SWPointer::scaled_iv_plus_offset: in(1) is offset_plus_k: ", n->in(1)->_idx); n->in(1)->dump();
3559 }
3560 }
3561
3562 void SWPointer::Tracer::scaled_iv_plus_offset_6(Node* n) {
3563 if(_slp->is_trace_alignment()) {
3564 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: Op_SubI PASSED", n->_idx);
3565 print_depth(); tty->print(" \\ %d SWPointer::scaled_iv_plus_offset: in(1) is scaled_iv: ", n->in(1)->_idx); n->in(1)->dump();
3566 print_depth(); tty->print(" \\ %d SWPointer::scaled_iv_plus_offset: in(2) is offset_plus_k: ", n->in(2)->_idx); n->in(2)->dump();
3567 }
3568 }
3569
3570 void SWPointer::Tracer::scaled_iv_plus_offset_7(Node* n) {
3571 if(_slp->is_trace_alignment()) {
3572 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: Op_SubI PASSED", n->_idx);
3573 print_depth(); tty->print(" \\ %d SWPointer::scaled_iv_plus_offset: in(2) is scaled_iv: ", n->in(2)->_idx); n->in(2)->dump();
3574 print_depth(); tty->print(" \\ %d SWPointer::scaled_iv_plus_offset: in(1) is offset_plus_k: ", n->in(1)->_idx); n->in(1)->dump();
3575 }
3576 }
3577
3578 void SWPointer::Tracer::scaled_iv_plus_offset_8(Node* n) {
3579 if(_slp->is_trace_alignment()) {
3580 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: FAILED", n->_idx);
3581 }
3582 }
3583
3584 void SWPointer::Tracer::scaled_iv_1(Node* n) {
3585 if(_slp->is_trace_alignment()) {
3586 print_depth(); tty->print(" %d SWPointer::scaled_iv: testing node: ", n->_idx); n->dump();
3587 }
3588 }
3589
3590 void SWPointer::Tracer::scaled_iv_2(Node* n, int scale) {
3591 if(_slp->is_trace_alignment()) {
3592 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: FAILED since another _scale has been detected before", n->_idx);
3593 print_depth(); tty->print_cr(" \\ SWPointer::scaled_iv: _scale (%d) != 0", scale);
3594 }
3595 }
3596
3597 void SWPointer::Tracer::scaled_iv_3(Node* n, int scale) {
3598 if(_slp->is_trace_alignment()) {
3599 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: is iv, setting _scale = %d", n->_idx, scale);
3600 }
3601 }
3602
3603 void SWPointer::Tracer::scaled_iv_4(Node* n, int scale) {
3604 if(_slp->is_trace_alignment()) {
3605 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: Op_MulI PASSED, setting _scale = %d", n->_idx, scale);
3606 print_depth(); tty->print(" \\ %d SWPointer::scaled_iv: in(1) is iv: ", n->in(1)->_idx); n->in(1)->dump();
3607 print_depth(); tty->print(" \\ %d SWPointer::scaled_iv: in(2) is Con: ", n->in(2)->_idx); n->in(2)->dump();
3608 }
3609 }
3610
3611 void SWPointer::Tracer::scaled_iv_5(Node* n, int scale) {
3612 if(_slp->is_trace_alignment()) {
3613 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: Op_MulI PASSED, setting _scale = %d", n->_idx, scale);
3614 print_depth(); tty->print(" \\ %d SWPointer::scaled_iv: in(2) is iv: ", n->in(2)->_idx); n->in(2)->dump();
3615 print_depth(); tty->print(" \\ %d SWPointer::scaled_iv: in(1) is Con: ", n->in(1)->_idx); n->in(1)->dump();
3616 }
3617 }
3618
3619 void SWPointer::Tracer::scaled_iv_6(Node* n, int scale) {
3620 if(_slp->is_trace_alignment()) {
3621 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: Op_LShiftI PASSED, setting _scale = %d", n->_idx, scale);
3622 print_depth(); tty->print(" \\ %d SWPointer::scaled_iv: in(1) is iv: ", n->in(1)->_idx); n->in(1)->dump();
3623 print_depth(); tty->print(" \\ %d SWPointer::scaled_iv: in(2) is Con: ", n->in(2)->_idx); n->in(2)->dump();
3624 }
3625 }
3626
3627 void SWPointer::Tracer::scaled_iv_7(Node* n) {
3628 if(_slp->is_trace_alignment()) {
3629 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: Op_ConvI2L PASSED", n->_idx);
3630 print_depth(); tty->print_cr(" \\ SWPointer::scaled_iv: in(1) %d is scaled_iv_plus_offset: ", n->in(1)->_idx);
3631 inc_depth(); inc_depth();
3632 print_depth(); n->in(1)->dump();
3633 dec_depth(); dec_depth();
3634 }
3635 }
3636
3637 void SWPointer::Tracer::scaled_iv_8(Node* n, SWPointer* tmp) {
3638 if(_slp->is_trace_alignment()) {
3639 print_depth(); tty->print(" %d SWPointer::scaled_iv: Op_LShiftL, creating tmp SWPointer: ", n->_idx); tmp->print();
3640 }
3641 }
3642
3643 void SWPointer::Tracer::scaled_iv_9(Node* n, int scale, int _offset, int mult) {
3644 if(_slp->is_trace_alignment()) {
3645 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: Op_LShiftL PASSED, setting _scale = %d, _offset = %d", n->_idx, scale, _offset);
3646 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",
3647 n->in(1)->_idx, n->in(2)->_idx, mult, scale, _offset);
3648 inc_depth(); inc_depth();
3649 print_depth(); n->in(1)->dump();
3650 print_depth(); n->in(2)->dump();
3651 dec_depth(); dec_depth();
3652 }
3653 }
3654
3655 void SWPointer::Tracer::scaled_iv_10(Node* n) {
3656 if(_slp->is_trace_alignment()) {
3657 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: FAILED", n->_idx);
3658 }
3659 }
3660
3661 void SWPointer::Tracer::offset_plus_k_1(Node* n) {
3662 if(_slp->is_trace_alignment()) {
3663 print_depth(); tty->print(" %d SWPointer::offset_plus_k: testing node: ", n->_idx); n->dump();
3664 }
3665 }
3666
3667 void SWPointer::Tracer::offset_plus_k_2(Node* n, int _offset) {
3668 if(_slp->is_trace_alignment()) {
3669 print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: Op_ConI PASSED, setting _offset = %d", n->_idx, _offset);
3670 }
3671 }
3672
3673 void SWPointer::Tracer::offset_plus_k_3(Node* n, int _offset) {
3674 if(_slp->is_trace_alignment()) {
3675 print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: Op_ConL PASSED, setting _offset = %d", n->_idx, _offset);
3676 }
3677 }
3678
3679 void SWPointer::Tracer::offset_plus_k_4(Node* n) {
3680 if(_slp->is_trace_alignment()) {
3681 print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: FAILED", n->_idx);
3682 print_depth(); tty->print_cr(" \\ " JLONG_FORMAT " SWPointer::offset_plus_k: Op_ConL FAILED, k is too big", n->get_long());
3683 }
3684 }
3685
3686 void SWPointer::Tracer::offset_plus_k_5(Node* n, Node* _invar) {
3687 if(_slp->is_trace_alignment()) {
3688 print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: FAILED since another invariant has been detected before", n->_idx);
3689 print_depth(); tty->print(" \\ %d SWPointer::offset_plus_k: _invar != NULL: ", _invar->_idx); _invar->dump();
3690 }
3691 }
3692
3693 void SWPointer::Tracer::offset_plus_k_6(Node* n, Node* _invar, bool _negate_invar, int _offset) {
3694 if(_slp->is_trace_alignment()) {
3695 print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: Op_AddI PASSED, setting _negate_invar = %d, _invar = %d, _offset = %d",
3696 n->_idx, _negate_invar, _invar->_idx, _offset);
3697 print_depth(); tty->print(" \\ %d SWPointer::offset_plus_k: in(2) is Con: ", n->in(2)->_idx); n->in(2)->dump();
3698 print_depth(); tty->print(" \\ %d SWPointer::offset_plus_k: in(1) is invariant: ", _invar->_idx); _invar->dump();
3699 }
3700 }
3701
3702 void SWPointer::Tracer::offset_plus_k_7(Node* n, Node* _invar, bool _negate_invar, int _offset) {
3703 if(_slp->is_trace_alignment()) {
3704 print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: Op_AddI PASSED, setting _negate_invar = %d, _invar = %d, _offset = %d",
3705 n->_idx, _negate_invar, _invar->_idx, _offset);
3706 print_depth(); tty->print(" \\ %d SWPointer::offset_plus_k: in(1) is Con: ", n->in(1)->_idx); n->in(1)->dump();
3707 print_depth(); tty->print(" \\ %d SWPointer::offset_plus_k: in(2) is invariant: ", _invar->_idx); _invar->dump();
3708 }
3709 }
3710
3711 void SWPointer::Tracer::offset_plus_k_8(Node* n, Node* _invar, bool _negate_invar, int _offset) {
3712 if(_slp->is_trace_alignment()) {
3713 print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: Op_SubI is PASSED, setting _negate_invar = %d, _invar = %d, _offset = %d",
3714 n->_idx, _negate_invar, _invar->_idx, _offset);
3715 print_depth(); tty->print(" \\ %d SWPointer::offset_plus_k: in(2) is Con: ", n->in(2)->_idx); n->in(2)->dump();
3716 print_depth(); tty->print(" \\ %d SWPointer::offset_plus_k: in(1) is invariant: ", _invar->_idx); _invar->dump();
3717 }
3718 }
3719
3720 void SWPointer::Tracer::offset_plus_k_9(Node* n, Node* _invar, bool _negate_invar, int _offset) {
3721 if(_slp->is_trace_alignment()) {
3722 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);
3723 print_depth(); tty->print(" \\ %d SWPointer::offset_plus_k: in(1) is Con: ", n->in(1)->_idx); n->in(1)->dump();
3724 print_depth(); tty->print(" \\ %d SWPointer::offset_plus_k: in(2) is invariant: ", _invar->_idx); _invar->dump();
3725 }
3726 }
3727
3728 void SWPointer::Tracer::offset_plus_k_10(Node* n, Node* _invar, bool _negate_invar, int _offset) {
3729 if(_slp->is_trace_alignment()) {
3730 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);
3731 print_depth(); tty->print_cr(" \\ %d SWPointer::offset_plus_k: is invariant", n->_idx);
3732 }
3733 }
3734
3735 void SWPointer::Tracer::offset_plus_k_11(Node* n) {
3736 if(_slp->is_trace_alignment()) {
3737 print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: FAILED", n->_idx);
3738 }
3739 }
3740
3741 #endif
3742 // ========================= OrderedPair =====================
3743
3744 const OrderedPair OrderedPair::initial;
3745
3746 // ========================= SWNodeInfo =====================
3747
3748 const SWNodeInfo SWNodeInfo::initial;
3749
3750
3751 // ============================ DepGraph ===========================
3752
3753 //------------------------------make_node---------------------------
3754 // Make a new dependence graph node for an ideal node.
3755 DepMem* DepGraph::make_node(Node* node) {
3756 DepMem* m = new (_arena) DepMem(node);
3757 if (node != NULL) {
3758 assert(_map.at_grow(node->_idx) == NULL, "one init only");
3759 _map.at_put_grow(node->_idx, m);
3760 }
3761 return m;
3762 }
3763
3764 //------------------------------make_edge---------------------------
3765 // Make a new dependence graph edge from dpred -> dsucc
3766 DepEdge* DepGraph::make_edge(DepMem* dpred, DepMem* dsucc) {
3767 DepEdge* e = new (_arena) DepEdge(dpred, dsucc, dsucc->in_head(), dpred->out_head());
3768 dpred->set_out_head(e);
3769 dsucc->set_in_head(e);
3770 return e;
3771 }
3772
3773 // ========================== DepMem ========================
3774
3775 //------------------------------in_cnt---------------------------
3776 int DepMem::in_cnt() {
3777 int ct = 0;
3778 for (DepEdge* e = _in_head; e != NULL; e = e->next_in()) ct++;
3779 return ct;
3780 }
3781
3782 //------------------------------out_cnt---------------------------
3783 int DepMem::out_cnt() {
3784 int ct = 0;
3785 for (DepEdge* e = _out_head; e != NULL; e = e->next_out()) ct++;
3786 return ct;
3787 }
3788
3789 //------------------------------print-----------------------------
3790 void DepMem::print() {
3791 #ifndef PRODUCT
3792 tty->print(" DepNode %d (", _node->_idx);
3793 for (DepEdge* p = _in_head; p != NULL; p = p->next_in()) {
3794 Node* pred = p->pred()->node();
3795 tty->print(" %d", pred != NULL ? pred->_idx : 0);
3796 }
3797 tty->print(") [");
3798 for (DepEdge* s = _out_head; s != NULL; s = s->next_out()) {
3799 Node* succ = s->succ()->node();
3800 tty->print(" %d", succ != NULL ? succ->_idx : 0);
3801 }
3802 tty->print_cr(" ]");
3803 #endif
3804 }
3805
3806 // =========================== DepEdge =========================
3807
3808 //------------------------------DepPreds---------------------------
3809 void DepEdge::print() {
3810 #ifndef PRODUCT
3811 tty->print_cr("DepEdge: %d [ %d ]", _pred->node()->_idx, _succ->node()->_idx);
3812 #endif
3813 }
3814
3815 // =========================== DepPreds =========================
3816 // Iterator over predecessor edges in the dependence graph.
3817
3818 //------------------------------DepPreds---------------------------
3819 DepPreds::DepPreds(Node* n, DepGraph& dg) {
3820 _n = n;
3821 _done = false;
3822 if (_n->is_Store() || _n->is_Load()) {
3823 _next_idx = MemNode::Address;
3824 _end_idx = n->req();
3825 _dep_next = dg.dep(_n)->in_head();
3826 } else if (_n->is_Mem()) {
3827 _next_idx = 0;
3828 _end_idx = 0;
3829 _dep_next = dg.dep(_n)->in_head();
3830 } else {
3831 _next_idx = 1;
3832 _end_idx = _n->req();
3833 _dep_next = NULL;
3834 }
3835 next();
3836 }
3837
3838 //------------------------------next---------------------------
3839 void DepPreds::next() {
3840 if (_dep_next != NULL) {
3841 _current = _dep_next->pred()->node();
3842 _dep_next = _dep_next->next_in();
3843 } else if (_next_idx < _end_idx) {
3844 _current = _n->in(_next_idx++);
3845 } else {
3846 _done = true;
3847 }
3848 }
3849
3850 // =========================== DepSuccs =========================
3851 // Iterator over successor edges in the dependence graph.
3852
3853 //------------------------------DepSuccs---------------------------
3854 DepSuccs::DepSuccs(Node* n, DepGraph& dg) {
3855 _n = n;
3856 _done = false;
3857 if (_n->is_Load()) {
3858 _next_idx = 0;
3859 _end_idx = _n->outcnt();
3860 _dep_next = dg.dep(_n)->out_head();
3861 } else if (_n->is_Mem() || _n->is_Phi() && _n->bottom_type() == Type::MEMORY) {
3862 _next_idx = 0;
3863 _end_idx = 0;
3864 _dep_next = dg.dep(_n)->out_head();
3865 } else {
3866 _next_idx = 0;
3867 _end_idx = _n->outcnt();
3868 _dep_next = NULL;
3869 }
3870 next();
3871 }
3872
3873 //-------------------------------next---------------------------
3874 void DepSuccs::next() {
3875 if (_dep_next != NULL) {
3876 _current = _dep_next->succ()->node();
3877 _dep_next = _dep_next->next_out();
3878 } else if (_next_idx < _end_idx) {
3879 _current = _n->raw_out(_next_idx++);
3880 } else {
3881 _done = true;
3882 }
3883 }
3884
3885 //
3886 // --------------------------------- vectorization/simd -----------------------------------
3887 //
3888 bool SuperWord::same_origin_idx(Node* a, Node* b) const {
3889 return a != NULL && b != NULL && _clone_map.same_idx(a->_idx, b->_idx);
3890 }
3891 bool SuperWord::same_generation(Node* a, Node* b) const {
3892 return a != NULL && b != NULL && _clone_map.same_gen(a->_idx, b->_idx);
3893 }
3894
3895 Node* SuperWord::find_phi_for_mem_dep(LoadNode* ld) {
3896 assert(in_bb(ld), "must be in block");
3897 if (_clone_map.gen(ld->_idx) == _ii_first) {
3898 #ifndef PRODUCT
3899 if (_vector_loop_debug) {
3900 tty->print_cr("SuperWord::find_phi_for_mem_dep _clone_map.gen(ld->_idx)=%d",
3901 _clone_map.gen(ld->_idx));
3902 }
3903 #endif
3904 return NULL; //we think that any ld in the first gen being vectorizable
3905 }
3906
3907 Node* mem = ld->in(MemNode::Memory);
3908 if (mem->outcnt() <= 1) {
3909 // we don't want to remove the only edge from mem node to load
3910 #ifndef PRODUCT
3911 if (_vector_loop_debug) {
3912 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",
3913 mem->_idx, ld->_idx);
3914 ld->dump();
3915 mem->dump();
3916 }
3917 #endif
3918 return NULL;
3919 }
3920 if (!in_bb(mem) || same_generation(mem, ld)) {
3921 #ifndef PRODUCT
3922 if (_vector_loop_debug) {
3923 tty->print_cr("SuperWord::find_phi_for_mem_dep _clone_map.gen(mem->_idx)=%d",
3924 _clone_map.gen(mem->_idx));
3925 }
3926 #endif
3927 return NULL; // does not depend on loop volatile node or depends on the same generation
3928 }
3929
3930 //otherwise first node should depend on mem-phi
3931 Node* first = first_node(ld);
3932 assert(first->is_Load(), "must be Load");
3933 Node* phi = first->as_Load()->in(MemNode::Memory);
3934 if (!phi->is_Phi() || phi->bottom_type() != Type::MEMORY) {
3935 #ifndef PRODUCT
3936 if (_vector_loop_debug) {
3937 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");
3938 ld->dump();
3939 first->dump();
3940 }
3941 #endif
3942 return NULL;
3943 }
3944
3945 Node* tail = 0;
3946 for (int m = 0; m < _mem_slice_head.length(); m++) {
3947 if (_mem_slice_head.at(m) == phi) {
3948 tail = _mem_slice_tail.at(m);
3949 }
3950 }
3951 if (tail == 0) { //test that found phi is in the list _mem_slice_head
3952 #ifndef PRODUCT
3953 if (_vector_loop_debug) {
3954 tty->print_cr("SuperWord::find_phi_for_mem_dep load %d is not vectorizable node, its phi %d is not _mem_slice_head",
3955 ld->_idx, phi->_idx);
3956 ld->dump();
3957 phi->dump();
3958 }
3959 #endif
3960 return NULL;
3961 }
3962
3963 // now all conditions are met
3964 return phi;
3965 }
3966
3967 Node* SuperWord::first_node(Node* nd) {
3968 for (int ii = 0; ii < _iteration_first.length(); ii++) {
3969 Node* nnn = _iteration_first.at(ii);
3970 if (same_origin_idx(nnn, nd)) {
3971 #ifndef PRODUCT
3972 if (_vector_loop_debug) {
3973 tty->print_cr("SuperWord::first_node: %d is the first iteration node for %d (_clone_map.idx(nnn->_idx) = %d)",
3974 nnn->_idx, nd->_idx, _clone_map.idx(nnn->_idx));
3975 }
3976 #endif
3977 return nnn;
3978 }
3979 }
3980
3981 #ifndef PRODUCT
3982 if (_vector_loop_debug) {
3983 tty->print_cr("SuperWord::first_node: did not find first iteration node for %d (_clone_map.idx(nd->_idx)=%d)",
3984 nd->_idx, _clone_map.idx(nd->_idx));
3985 }
3986 #endif
3987 return 0;
3988 }
3989
3990 Node* SuperWord::last_node(Node* nd) {
3991 for (int ii = 0; ii < _iteration_last.length(); ii++) {
3992 Node* nnn = _iteration_last.at(ii);
3993 if (same_origin_idx(nnn, nd)) {
3994 #ifndef PRODUCT
3995 if (_vector_loop_debug) {
3996 tty->print_cr("SuperWord::last_node _clone_map.idx(nnn->_idx)=%d, _clone_map.idx(nd->_idx)=%d",
3997 _clone_map.idx(nnn->_idx), _clone_map.idx(nd->_idx));
3998 }
3999 #endif
4000 return nnn;
4001 }
4002 }
4003 return 0;
4004 }
4005
4006 int SuperWord::mark_generations() {
4007 Node *ii_err = 0, *tail_err;
4008 for (int i = 0; i < _mem_slice_head.length(); i++) {
4009 Node* phi = _mem_slice_head.at(i);
4010 assert(phi->is_Phi(), "must be phi");
4011
4012 Node* tail = _mem_slice_tail.at(i);
4013 if (_ii_last == -1) {
4014 tail_err = tail;
4015 _ii_last = _clone_map.gen(tail->_idx);
4016 }
4017 else if (_ii_last != _clone_map.gen(tail->_idx)) {
4018 #ifndef PRODUCT
4019 if (TraceSuperWord && Verbose) {
4020 tty->print_cr("SuperWord::mark_generations _ii_last error - found different generations in two tail nodes ");
4021 tail->dump();
4022 tail_err->dump();
4023 }
4024 #endif
4025 return -1;
4026 }
4027
4028 // find first iteration in the loop
4029 for (DUIterator_Fast imax, i = phi->fast_outs(imax); i < imax; i++) {
4030 Node* ii = phi->fast_out(i);
4031 if (in_bb(ii) && ii->is_Store()) { // we speculate that normally Stores of one and one only generation have deps from mem phi
4032 if (_ii_first == -1) {
4033 ii_err = ii;
4034 _ii_first = _clone_map.gen(ii->_idx);
4035 } else if (_ii_first != _clone_map.gen(ii->_idx)) {
4036 #ifndef PRODUCT
4037 if (TraceSuperWord && Verbose) {
4038 tty->print_cr("SuperWord::mark_generations: _ii_first was found before and not equal to one in this node (%d)", _ii_first);
4039 ii->dump();
4040 if (ii_err!= 0) {
4041 ii_err->dump();
4042 }
4043 }
4044 #endif
4045 return -1; // this phi has Stores from different generations of unroll and cannot be simd/vectorized
4046 }
4047 }
4048 }//for (DUIterator_Fast imax,
4049 }//for (int i...
4050
4051 if (_ii_first == -1 || _ii_last == -1) {
4052 #ifndef PRODUCT
4053 if (TraceSuperWord && Verbose) {
4054 tty->print_cr("SuperWord::mark_generations unknown error, something vent wrong");
4055 }
4056 #endif
4057 return -1; // something vent wrong
4058 }
4059 // collect nodes in the first and last generations
4060 assert(_iteration_first.length() == 0, "_iteration_first must be empty");
4061 assert(_iteration_last.length() == 0, "_iteration_last must be empty");
4062 for (int j = 0; j < _block.length(); j++) {
4063 Node* n = _block.at(j);
4064 node_idx_t gen = _clone_map.gen(n->_idx);
4065 if ((signed)gen == _ii_first) {
4066 _iteration_first.push(n);
4067 } else if ((signed)gen == _ii_last) {
4068 _iteration_last.push(n);
4069 }
4070 }
4071
4072 // building order of iterations
4073 if (_ii_order.length() == 0 && ii_err != 0) {
4074 assert(in_bb(ii_err) && ii_err->is_Store(), "should be Store in bb");
4075 Node* nd = ii_err;
4076 while(_clone_map.gen(nd->_idx) != _ii_last) {
4077 _ii_order.push(_clone_map.gen(nd->_idx));
4078 bool found = false;
4079 for (DUIterator_Fast imax, i = nd->fast_outs(imax); i < imax; i++) {
4080 Node* use = nd->fast_out(i);
4081 if (same_origin_idx(use, nd) && use->as_Store()->in(MemNode::Memory) == nd) {
4082 found = true;
4083 nd = use;
4084 break;
4085 }
4086 }//for
4087
4088 if (found == false) {
4089 #ifndef PRODUCT
4090 if (TraceSuperWord && Verbose) {
4091 tty->print_cr("SuperWord::mark_generations: Cannot build order of iterations - no dependent Store for %d", nd->_idx);
4092 }
4093 #endif
4094 _ii_order.clear();
4095 return -1;
4096 }
4097 } //while
4098 _ii_order.push(_clone_map.gen(nd->_idx));
4099 }
4100
4101 #ifndef PRODUCT
4102 if (_vector_loop_debug) {
4103 tty->print_cr("SuperWord::mark_generations");
4104 tty->print_cr("First generation (%d) nodes:", _ii_first);
4105 for (int ii = 0; ii < _iteration_first.length(); ii++) _iteration_first.at(ii)->dump();
4106 tty->print_cr("Last generation (%d) nodes:", _ii_last);
4107 for (int ii = 0; ii < _iteration_last.length(); ii++) _iteration_last.at(ii)->dump();
4108 tty->print_cr(" ");
4109
4110 tty->print("SuperWord::List of generations: ");
4111 for (int jj = 0; jj < _ii_order.length(); ++jj) {
4112 tty->print("%d:%d ", jj, _ii_order.at(jj));
4113 }
4114 tty->print_cr(" ");
4115 }
4116 #endif
4117
4118 return _ii_first;
4119 }
4120
4121 bool SuperWord::fix_commutative_inputs(Node* gold, Node* fix) {
4122 assert(gold->is_Add() && fix->is_Add() || gold->is_Mul() && fix->is_Mul(), "should be only Add or Mul nodes");
4123 assert(same_origin_idx(gold, fix), "should be clones of the same node");
4124 Node* gin1 = gold->in(1);
4125 Node* gin2 = gold->in(2);
4126 Node* fin1 = fix->in(1);
4127 Node* fin2 = fix->in(2);
4128 bool swapped = false;
4129
4130 if (in_bb(gin1) && in_bb(gin2) && in_bb(fin1) && in_bb(fin1)) {
4131 if (same_origin_idx(gin1, fin1) &&
4132 same_origin_idx(gin2, fin2)) {
4133 return true; // nothing to fix
4134 }
4135 if (same_origin_idx(gin1, fin2) &&
4136 same_origin_idx(gin2, fin1)) {
4137 fix->swap_edges(1, 2);
4138 swapped = true;
4139 }
4140 }
4141 // at least one input comes from outside of bb
4142 if (gin1->_idx == fin1->_idx) {
4143 return true; // nothing to fix
4144 }
4145 if (!swapped && (gin1->_idx == fin2->_idx || gin2->_idx == fin1->_idx)) { //swapping is expensive, check condition first
4146 fix->swap_edges(1, 2);
4147 swapped = true;
4148 }
4149
4150 if (swapped) {
4151 #ifndef PRODUCT
4152 if (_vector_loop_debug) {
4153 tty->print_cr("SuperWord::fix_commutative_inputs: fixed node %d", fix->_idx);
4154 }
4155 #endif
4156 return true;
4157 }
4158
4159 #ifndef PRODUCT
4160 if (TraceSuperWord && Verbose) {
4161 tty->print_cr("SuperWord::fix_commutative_inputs: cannot fix node %d", fix->_idx);
4162 }
4163 #endif
4164 return false;
4165 }
4166
4167 bool SuperWord::pack_parallel() {
4168 #ifndef PRODUCT
4169 if (_vector_loop_debug) {
4170 tty->print_cr("SuperWord::pack_parallel: START");
4171 }
4172 #endif
4173
4174 _packset.clear();
4175
4176 for (int ii = 0; ii < _iteration_first.length(); ii++) {
4177 Node* nd = _iteration_first.at(ii);
4178 if (in_bb(nd) && (nd->is_Load() || nd->is_Store() || nd->is_Add() || nd->is_Mul())) {
4179 Node_List* pk = new Node_List();
4180 pk->push(nd);
4181 for (int gen = 1; gen < _ii_order.length(); ++gen) {
4182 for (int kk = 0; kk < _block.length(); kk++) {
4183 Node* clone = _block.at(kk);
4184 if (same_origin_idx(clone, nd) &&
4185 _clone_map.gen(clone->_idx) == _ii_order.at(gen)) {
4186 if (nd->is_Add() || nd->is_Mul()) {
4187 fix_commutative_inputs(nd, clone);
4188 }
4189 pk->push(clone);
4190 if (pk->size() == 4) {
4191 _packset.append(pk);
4192 #ifndef PRODUCT
4193 if (_vector_loop_debug) {
4194 tty->print_cr("SuperWord::pack_parallel: added pack ");
4195 pk->dump();
4196 }
4197 #endif
4198 if (_clone_map.gen(clone->_idx) != _ii_last) {
4199 pk = new Node_List();
4200 }
4201 }
4202 break;
4203 }
4204 }
4205 }//for
4206 }//if
4207 }//for
4208
4209 #ifndef PRODUCT
4210 if (_vector_loop_debug) {
4211 tty->print_cr("SuperWord::pack_parallel: END");
4212 }
4213 #endif
4214
4215 return true;
4216 }
4217
4218 bool SuperWord::hoist_loads_in_graph() {
4219 GrowableArray<Node*> loads;
4220
4221 #ifndef PRODUCT
4222 if (_vector_loop_debug) {
4223 tty->print_cr("SuperWord::hoist_loads_in_graph: total number _mem_slice_head.length() = %d", _mem_slice_head.length());
4224 }
4225 #endif
4226
4227 for (int i = 0; i < _mem_slice_head.length(); i++) {
4228 Node* n = _mem_slice_head.at(i);
4229 if ( !in_bb(n) || !n->is_Phi() || n->bottom_type() != Type::MEMORY) {
4230 #ifndef PRODUCT
4231 if (TraceSuperWord && Verbose) {
4232 tty->print_cr("SuperWord::hoist_loads_in_graph: skipping unexpected node n=%d", n->_idx);
4233 }
4234 #endif
4235 continue;
4236 }
4237
4238 #ifndef PRODUCT
4239 if (_vector_loop_debug) {
4240 tty->print_cr("SuperWord::hoist_loads_in_graph: processing phi %d = _mem_slice_head.at(%d);", n->_idx, i);
4241 }
4242 #endif
4243
4244 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4245 Node* ld = n->fast_out(i);
4246 if (ld->is_Load() && ld->as_Load()->in(MemNode::Memory) == n && in_bb(ld)) {
4247 for (int i = 0; i < _block.length(); i++) {
4248 Node* ld2 = _block.at(i);
4249 if (ld2->is_Load() && same_origin_idx(ld, ld2) &&
4250 !same_generation(ld, ld2)) { // <= do not collect the first generation ld
4251 #ifndef PRODUCT
4252 if (_vector_loop_debug) {
4253 tty->print_cr("SuperWord::hoist_loads_in_graph: will try to hoist load ld2->_idx=%d, cloned from %d (ld->_idx=%d)",
4254 ld2->_idx, _clone_map.idx(ld->_idx), ld->_idx);
4255 }
4256 #endif
4257 // could not do on-the-fly, since iterator is immutable
4258 loads.push(ld2);
4259 }
4260 }// for
4261 }//if
4262 }//for (DUIterator_Fast imax,
4263 }//for (int i = 0; i
4264
4265 for (int i = 0; i < loads.length(); i++) {
4266 LoadNode* ld = loads.at(i)->as_Load();
4267 Node* phi = find_phi_for_mem_dep(ld);
4268 if (phi != NULL) {
4269 #ifndef PRODUCT
4270 if (_vector_loop_debug) {
4271 tty->print_cr("SuperWord::hoist_loads_in_graph replacing MemNode::Memory(%d) edge in %d with one from %d",
4272 MemNode::Memory, ld->_idx, phi->_idx);
4273 }
4274 #endif
4275 _igvn.replace_input_of(ld, MemNode::Memory, phi);
4276 }
4277 }//for
4278
4279 restart(); // invalidate all basic structures, since we rebuilt the graph
4280
4281 #ifndef PRODUCT
4282 if (TraceSuperWord && Verbose) {
4283 tty->print_cr("\nSuperWord::hoist_loads_in_graph() the graph was rebuilt, all structures invalidated and need rebuild");
4284 }
4285 #endif
4286 return true;
4287 }
4288